agora inbox for [email protected]help / 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 <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 @ 2025-07-26 00:57 Lukas Fittl <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Lukas Fittl @ 2025-07-26 00:57 UTC (permalink / raw) We switch to using the time stamp counter (TSC) instead of clock_gettime() to reduce overhead of EXPLAIN (ANALYZE, TIME ON). Tests showed that runtime is reduced by around 10% for queries moving lots of rows through the plan. For now this is only enabled on Linux/x86, in case the system clocksource is reported as TSC. Relying on the Linux kernel simplifies the logic to detect if the present TSC is usable (frequency invariant, synchronized between sockets, etc.). In all other cases we fallback to clock_gettime(). Note, that we intentionally use RDTSC in the fast paths, rather than RDTSCP. RDTSCP waits for outstanding instructions to retire on out-of-order CPUs. This adds noticably for little benefit in the typical InstrStartNode() / InstrStopNode() use case. The macro to be used in such cases is called INSTR_TIME_SET_CURRENT_FAST(). The original macro INSTR_TIME_SET_CURRENT() uses RDTSCP and is supposed to be used when precision is more important than performance. Author: David Geier <[email protected]> Author: Andres Freund <[email protected]> Author: Lukas Fittl <[email protected]> Reviewed-by: Discussion: https://www.postgresql.org/message-id/flat/20200612232810.f46nbqkdhbutzqdg%40alap3.anarazel.de --- src/backend/access/heap/vacuumlazy.c | 4 +- src/backend/executor/instrument.c | 12 +- src/backend/utils/init/postinit.c | 3 + src/bin/pgbench/pgbench.c | 3 + src/bin/psql/startup.c | 4 + src/common/Makefile | 1 + src/common/instr_time.c | 206 +++++++++++++++++++++++++++ src/common/meson.build | 1 + src/include/portability/instr_time.h | 136 +++++++++++++++--- 9 files changed, 348 insertions(+), 22 deletions(-) create mode 100644 src/common/instr_time.c diff --git a/src/backend/access/heap/vacuumlazy.c b/src/backend/access/heap/vacuumlazy.c index d2b031fdd06..5027048cac4 100644 --- a/src/backend/access/heap/vacuumlazy.c +++ b/src/backend/access/heap/vacuumlazy.c @@ -3409,8 +3409,8 @@ count_nondeletable_pages(LVRelState *vacrel, bool *lock_waiter_detected) INSTR_TIME_SET_CURRENT(currenttime); elapsed = currenttime; INSTR_TIME_SUBTRACT(elapsed, starttime); - if ((INSTR_TIME_GET_MICROSEC(elapsed) / 1000) - >= VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) + if (INSTR_TIME_GET_MILLISEC(elapsed) >= + VACUUM_TRUNCATE_LOCK_CHECK_INTERVAL) { if (LockHasWaitersRelation(vacrel->rel, AccessExclusiveLock)) { diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c index 56e635f4700..01f67c5d972 100644 --- a/src/backend/executor/instrument.c +++ b/src/backend/executor/instrument.c @@ -67,9 +67,13 @@ InstrInit(Instrumentation *instr, int instrument_options) void InstrStartNode(Instrumentation *instr) { - if (instr->need_timer && - !INSTR_TIME_SET_CURRENT_LAZY(instr->starttime)) - elog(ERROR, "InstrStartNode called twice in a row"); + if (instr->need_timer) + { + if (!INSTR_TIME_IS_ZERO(instr->starttime)) + elog(ERROR, "InstrStartNode called twice in a row"); + else + INSTR_TIME_SET_CURRENT_FAST(instr->starttime); + } /* save buffer usage totals at node entry, if needed */ if (instr->need_bufusage) @@ -95,7 +99,7 @@ InstrStopNode(Instrumentation *instr, double nTuples) if (INSTR_TIME_IS_ZERO(instr->starttime)) elog(ERROR, "InstrStopNode called without start"); - INSTR_TIME_SET_CURRENT(endtime); + INSTR_TIME_SET_CURRENT_FAST(endtime); INSTR_TIME_ACCUM_DIFF(instr->counter, endtime, instr->starttime); INSTR_TIME_SET_ZERO(instr->starttime); diff --git a/src/backend/utils/init/postinit.c b/src/backend/utils/init/postinit.c index 641e535a73c..d573409903b 100644 --- a/src/backend/utils/init/postinit.c +++ b/src/backend/utils/init/postinit.c @@ -810,6 +810,9 @@ InitPostgres(const char *in_dbname, Oid dboid, /* Initialize portal manager */ EnablePortalManager(); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* * Load relcache entries for the shared system catalogs. This must create * at least entries for pg_database and catalogs used for authentication. diff --git a/src/bin/pgbench/pgbench.c b/src/bin/pgbench/pgbench.c index 1515ed405ba..79bef2d2aec 100644 --- a/src/bin/pgbench/pgbench.c +++ b/src/bin/pgbench/pgbench.c @@ -7290,6 +7290,9 @@ main(int argc, char **argv) initRandomState(&state[i].cs_func_rs); } + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + /* opening connection... */ con = doConnect(); if (con == NULL) diff --git a/src/bin/psql/startup.c b/src/bin/psql/startup.c index 249b6aa5169..d615df593c7 100644 --- a/src/bin/psql/startup.c +++ b/src/bin/psql/startup.c @@ -24,6 +24,7 @@ #include "help.h" #include "input.h" #include "mainloop.h" +#include "portability/instr_time.h" #include "settings.h" /* @@ -327,6 +328,9 @@ main(int argc, char *argv[]) PQsetNoticeProcessor(pset.db, NoticeProcessor, NULL); + /* initialize high-precision interval timing */ + INSTR_TIME_INITIALIZE(); + SyncVariables(); if (options.list_dbs) diff --git a/src/common/Makefile b/src/common/Makefile index 2c720caa509..1a2fbbe887f 100644 --- a/src/common/Makefile +++ b/src/common/Makefile @@ -59,6 +59,7 @@ OBJS_COMMON = \ file_perm.o \ file_utils.o \ hashfn.o \ + instr_time.o \ ip.o \ jsonapi.o \ keywords.o \ diff --git a/src/common/instr_time.c b/src/common/instr_time.c new file mode 100644 index 00000000000..fdf47699f20 --- /dev/null +++ b/src/common/instr_time.c @@ -0,0 +1,206 @@ +/*------------------------------------------------------------------------- + * + * instr_time.c + * Non-inline parts of the portable high-precision interval timing + * implementation + * + * Portions Copyright (c) 2025, PostgreSQL Global Development Group + * + * + * IDENTIFICATION + * src/backend/port/instr_time.c + * + *------------------------------------------------------------------------- + */ +#include "postgres.h" + +#if defined(HAVE__GET_CPUID) || (defined(HAVE__CPUIDEX) && !defined(_MSC_VER)) +#include <cpuid.h> +#endif + +#if defined(HAVE__CPUID) || (defined(HAVE__CPUIDEX) && defined(_MSC_VER)) +#include <intrin.h> +#endif + +#include "portability/instr_time.h" + +#ifndef WIN32 +/* + * Stores what the number of cycles needs to be multiplied with to end up + * with nanoseconds using integer math. See comment in pg_initialize_rdtsc() + * for more details. + * + * By default assume we are using clock_gettime() as a fallback which uses + * nanoseconds as ticks. Hence, we set the multiplier to the precision scalar + * so that the division in INSTR_TIME_GET_NANOSEC() won't change the nanoseconds. + * + * When using the RDTSC instruction directly this is filled in during initialization + * based on the relevant CPUID fields. + */ +int64 ticks_per_ns_scaled = TICKS_TO_NS_PRECISION; +int64 ticks_per_sec = NS_PER_S; +int64 max_ticks_no_overflow = PG_INT64_MAX / TICKS_TO_NS_PRECISION; + +#if defined(__x86_64__) && defined(__linux__) +/* + * Indicates if RDTSC can be used (Linux/x86 only, when OS uses TSC clocksource) + */ +bool has_rdtsc = false; + +/* + * Indicates if RDTSCP can be used. True if RDTSC can be used and RDTSCP is available. + */ +bool has_rdtscp = false; + +#define CPUID_HYPERVISOR_VMWARE(words) (words[1] == 0x61774d56 && words[2] == 0x4d566572 && words[3] == 0x65726177) /* VMwareVMware */ +#define CPUID_HYPERVISOR_KVM(words) (words[1] == 0x4b4d564b && words[2] == 0x564b4d56 && words[3] == 0x0000004d) /* KVMKVMKVM */ + +static bool +get_tsc_frequency_khz(uint32 *tsc_freq) +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x15, &r[0] /* denominator */ , &r[1] /* numerator */ , &r[2] /* hz */ , &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x15); +#else +#error cpuid instruction not available +#endif + + if (r[2] > 0) + { + if (r[0] == 0 || r[1] == 0) + return false; + + *tsc_freq = r[2] / 1000 * r[1] / r[0]; + return true; + } + + /* Some CPUs only report frequency in 16H */ + +#if defined(HAVE__GET_CPUID) + __get_cpuid(0x16, &r[0] /* base_mhz */ , &r[1], &r[2], &r[3]); +#elif defined(HAVE__CPUID) + __cpuid(r, 0x16); +#else +#error cpuid instruction not available +#endif + + if (r[0] > 0) + { + *tsc_freq = r[0] * 1000; + return true; + } + + /* + * Check if we have a KVM or VMware Hypervisor passing down TSC frequency + * to us in a guest VM + * + * Note that accessing the 0x40000000 leaf for Hypervisor info requires + * use of __cpuidex to set ECX to 0. The similar __get_cpuid_count + * function does not work as expected since it contains a check for + * __get_cpuid_max, which has been observed to be lower than the special + * Hypervisor leaf. + */ +#if defined(HAVE__CPUIDEX) + __cpuidex((int32 *) r, 0x40000000, 0); + if (r[0] >= 0x40000010 && (CPUID_HYPERVISOR_VMWARE(r) || CPUID_HYPERVISOR_KVM(r))) + { + __cpuidex((int32 *) r, 0x40000010, 0); + if (r[0] > 0) + { + *tsc_freq = r[0]; + return true; + } + } +#endif + + return false; +} + +static bool +is_rdtscp_available() +{ + uint32 r[4] = {0, 0, 0, 0}; + +#if defined(HAVE__GET_CPUID) + if (!__get_cpuid(0x80000001, &r[0], &r[1], &r[2], &r[3])) + return false; +#elif defined(HAVE__CPUID) + __cpuid(r, 0x80000001); +#else +#error cpuid instruction not available +#endif + + return (r[3] & (1 << 27)) != 0; +} + +/* + * Decide whether we use the RDTSC instruction at runtime, for Linux/x86, + * instead of incurring the overhead of a full clock_gettime() call. + * + * This can't be reliably determined at compile time, since the + * availability of an "invariant" TSC (that is not affected by CPU + * frequency changes) is dependent on the CPU architecture. Additionally, + * there are cases where TSC availability is impacted by virtualization, + * where a simple cpuid feature check would not be enough. + * + * Since Linux already does a significant amount of work to determine + * whether TSC is a viable clock source, decide based on that. + */ +void +pg_initialize_rdtsc(void) +{ + FILE *fp = fopen("/sys/devices/system/clocksource/clocksource0/current_clocksource", "r"); + + if (fp) + { + char buf[128]; + + if (fgets(buf, sizeof(buf), fp) != NULL && strcmp(buf, "tsc\n") == 0) + { + /* + * Compute baseline CPU peformance, determines speed at which + * RDTSC advances. + */ + uint32 tsc_freq; + + if (get_tsc_frequency_khz(&tsc_freq)) + { + /* + * Ticks to nanoseconds conversion requires floating point + * math because because: + * + * sec = ticks / frequency_hz ns = ticks / frequency_hz * + * 1,000,000,000 ns = ticks * (1,000,000,000 / frequency_hz) + * ns = ticks * (1,000,000 / frequency_khz) <-- now in + * kilohertz + * + * Here, 'ns' is usually a floating number. For example for a + * 2.5 GHz CPU the scaling factor becomes 1,000,000 / + * 2,500,000 = 1.2. + * + * To be able to use integer math we work around the lack of + * precision. We first scale the integer up and after the + * multiplication by the number of ticks in + * INSTR_TIME_GET_NANOSEC() we divide again by the same value. + * We picked the scaler such that it provides enough precision + * and is a power-of-two which allows for shifting instead of + * doing an integer division. + */ + ticks_per_ns_scaled = INT64CONST(1000000) * TICKS_TO_NS_PRECISION / tsc_freq; + ticks_per_sec = tsc_freq * 1000; /* KHz->Hz */ + max_ticks_no_overflow = PG_INT64_MAX / ticks_per_ns_scaled; + + has_rdtsc = true; + has_rdtscp = is_rdtscp_available(); + } + } + + fclose(fp); + } +} +#endif /* defined(__x86_64__) && defined(__linux__) */ + +#endif /* WIN32 */ diff --git a/src/common/meson.build b/src/common/meson.build index 1540ba67cca..62b90b3e609 100644 --- a/src/common/meson.build +++ b/src/common/meson.build @@ -13,6 +13,7 @@ common_sources = files( 'file_perm.c', 'file_utils.c', 'hashfn.c', + 'instr_time.c', 'ip.c', 'jsonapi.c', 'keywords.c', diff --git a/src/include/portability/instr_time.h b/src/include/portability/instr_time.h index f71a851b18d..e2e339a0c4f 100644 --- a/src/include/portability/instr_time.h +++ b/src/include/portability/instr_time.h @@ -4,9 +4,11 @@ * portable high-precision interval timing * * This file provides an abstraction layer to hide portability issues in - * interval timing. On Unix we use clock_gettime(), and on Windows we use - * QueryPerformanceCounter(). These macros also give some breathing room to - * use other high-precision-timing APIs. + * interval timing. On Linux/x86 we use the rdtsc instruction when a TSC + * clocksource is also used on the host OS. Otherwise, and on other + * Unix-like systems we use clock_gettime() and on Windows we use + * QueryPerformanceCounter(). These macros also give some breathing + * room to use other high-precision-timing APIs. * * The basic data type is instr_time, which all callers should treat as an * opaque typedef. instr_time can store either an absolute time (of @@ -17,10 +19,11 @@ * * INSTR_TIME_SET_ZERO(t) set t to zero (memset is acceptable too) * - * INSTR_TIME_SET_CURRENT(t) set t to current time + * INSTR_TIME_SET_CURRENT_FAST(t) set t to current time without waiting + * for instructions in out-of-order window * - * INSTR_TIME_SET_CURRENT_LAZY(t) set t to current time if t is zero, - * evaluates to whether t changed + * INSTR_TIME_SET_CURRENT(t) set t to current time while waiting for + * instructions in OOO to retire * * INSTR_TIME_ADD(x, y) x += y * @@ -81,6 +84,15 @@ typedef struct instr_time #ifndef WIN32 +/* + * Make sure this is a power-of-two, so that the compiler can turn the + * multiplications and divisions into shifts. + */ +#define TICKS_TO_NS_PRECISION (1<<14) + +extern int64 ticks_per_ns_scaled; +extern int64 ticks_per_sec; +extern int64 max_ticks_no_overflow; /* Use clock_gettime() */ @@ -106,9 +118,18 @@ typedef struct instr_time #define PG_INSTR_CLOCK CLOCK_REALTIME #endif -/* helper for INSTR_TIME_SET_CURRENT */ +#if defined(__x86_64__) && defined(__linux__) +#include <x86intrin.h> +#include <cpuid.h> + +extern bool has_rdtsc; +extern bool has_rdtscp; + +extern void pg_initialize_rdtsc(void); +#endif + static inline instr_time -pg_clock_gettime_ns(void) +pg_clock_gettime(void) { instr_time now; struct timespec tmp; @@ -119,11 +140,94 @@ pg_clock_gettime_ns(void) return now; } +static inline instr_time +pg_get_ticks_fast(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtsc) + { + instr_time now; + + now.ticks = __rdtsc(); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline instr_time +pg_get_ticks(void) +{ +#if defined(__x86_64__) && defined(__linux__) + if (has_rdtscp) + { + instr_time now; + uint32 unused; + + now.ticks = __rdtscp(&unused); + return now; + } +#endif + + return pg_clock_gettime(); +} + +static inline int64_t +pg_ticks_to_ns(instr_time t) +{ + /* + * Would multiplication overflow? If so perform computation in two parts. + * Check overflow without actually overflowing via: a * b > max <=> a > + * max / b + */ + int64 ns = 0; + + if (unlikely(t.ticks > max_ticks_no_overflow)) + { + /* + * Compute how often the maximum number of ticks fits completely into + * the number of elapsed ticks and convert that number into + * nanoseconds. Then multiply by the count to arrive at the final + * value. In a 2nd step we adjust the number of elapsed ticks and + * convert the remaining ticks. + */ + int64 count = t.ticks / max_ticks_no_overflow; + int64 max_ns = max_ticks_no_overflow * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + + ns = max_ns * count; + + /* + * Subtract the ticks that we now already accounted for, so that they + * don't get counted twice. + */ + t.ticks -= count * max_ticks_no_overflow; + Assert(t.ticks >= 0); + } + + ns += t.ticks * ticks_per_ns_scaled / TICKS_TO_NS_PRECISION; + return ns; +} + +static inline void +pg_initialize_get_ticks() +{ +#if defined(__x86_64__) && defined(__linux__) + pg_initialize_rdtsc(); +#endif +} + +#define INSTR_TIME_INITIALIZE() \ + pg_initialize_get_ticks() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_get_ticks_fast()) + #define INSTR_TIME_SET_CURRENT(t) \ - ((t) = pg_clock_gettime_ns()) + ((t) = pg_get_ticks()) #define INSTR_TIME_GET_NANOSEC(t) \ - ((int64) (t).ticks) + pg_ticks_to_ns(t) #else /* WIN32 */ @@ -131,7 +235,7 @@ pg_clock_gettime_ns(void) /* Use QueryPerformanceCounter() */ -/* helper for INSTR_TIME_SET_CURRENT */ +/* helper for INSTR_TIME_SET_CURRENT / INSTR_TIME_SET_CURRENT_FAST */ static inline instr_time pg_query_performance_counter(void) { @@ -153,6 +257,11 @@ GetTimerFrequency(void) return (double) f.QuadPart; } +#define INSTR_TIME_INITIALIZE() + +#define INSTR_TIME_SET_CURRENT_FAST(t) \ + ((t) = pg_query_performance_counter()) + #define INSTR_TIME_SET_CURRENT(t) \ ((t) = pg_query_performance_counter()) @@ -168,13 +277,8 @@ GetTimerFrequency(void) #define INSTR_TIME_IS_ZERO(t) ((t).ticks == 0) - #define INSTR_TIME_SET_ZERO(t) ((t).ticks = 0) -#define INSTR_TIME_SET_CURRENT_LAZY(t) \ - (INSTR_TIME_IS_ZERO(t) ? INSTR_TIME_SET_CURRENT(t), true : false) - - #define INSTR_TIME_ADD(x,y) \ ((x).ticks += (y).ticks) -- 2.47.3 --vtqqtrpooseurzip Content-Type: text/x-diff; charset=utf-8 Content-Disposition: attachment; filename="v12-0003-pg_test_timing-Add-fast-flag-to-test-fast-timing.patch" ^ permalink raw reply [nested|flat] 268+ messages in thread
* [PATCH 7/8] Decouple updating of freezing information from swap_relation_files(). @ 2026-07-15 08:37 Antonin Houska <[email protected]> 0 siblings, 0 replies; 268+ messages in thread From: Antonin Houska @ 2026-07-15 08:37 UTC (permalink / raw) So far, the relfrozenxid and relminmxid attributes of pg_class used to be updated by swap_relation_files(), however, not every call of the function should do that. On the contrary, new feature of REPACK (CONCURRENTLY) will need to update this information in one case that would be tricky to handle using swap_relation_files(). --- src/backend/commands/repack.c | 114 +++++++++++++++++----------------- 1 file changed, 56 insertions(+), 58 deletions(-) diff --git a/src/backend/commands/repack.c b/src/backend/commands/repack.c index 89695335127..7390fd303f6 100644 --- a/src/backend/commands/repack.c +++ b/src/backend/commands/repack.c @@ -188,6 +188,8 @@ static void copy_table_data(Relation NewHeap, Relation OldHeap, Relation OldInde TransactionId *pFreezeXid, MultiXactId *pCutoffMulti, ChangeContext *chgcxt); +static void update_relation_cutoffs(Oid relid, TransactionId frozenXid, + MultiXactId cutoffMulti); static List *get_tables_to_repack(RepackCommand cmd, bool usingindex, MemoryContext permcxt); static List *get_tables_to_repack_partitioned(RepackCommand cmd, @@ -2039,13 +2041,6 @@ copy_table_data(Relation NewHeap, Relation OldHeap, Relation OldIndex, * links) while the latter is the only way to handle cases in which a toast * table is added or removed altogether. * - * Additionally, the first relation is marked with relfrozenxid set to - * frozenXid. It seems a bit ugly to have this here, but the caller would - * have to do it anyway, so having it here saves a heap_update. Note: in - * the swap-toast-links case, we assume we don't need to change the toast - * table's relfrozenxid: the new version of the toast table should already - * have relfrozenxid set to RecentXmin, which is good enough. - * * Lastly, if r2 and its toast table and toast index (if any) are mapped, * their OIDs are emitted into mapped_tables[]. This is hacky but beats * having to look the information up again later in finish_heap_swap. @@ -2054,9 +2049,7 @@ static void swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, bool swap_toast_by_content, bool is_internal, - TransactionId frozenXid, - MultiXactId cutoffMulti, - Oid *mapped_tables) + Oid *mapped_tables, Oid *r1_toastid) { Relation relRelation; HeapTuple reltup1, @@ -2088,6 +2081,9 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, relam1 = relform1->relam; relam2 = relform2->relam; + if (r1_toastid) + *r1_toastid = relform1->reltoastrelid; + if (RelFileNumberIsValid(relfilenumber1) && RelFileNumberIsValid(relfilenumber2)) { @@ -2203,15 +2199,6 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, * and then fail to commit the pg_class update. */ - /* set rel1's frozen Xid and minimum MultiXid */ - if (relform1->relkind != RELKIND_INDEX) - { - Assert(!TransactionIdIsValid(frozenXid) || - TransactionIdIsNormal(frozenXid)); - relform1->relfrozenxid = frozenXid; - relform1->relminmxid = cutoffMulti; - } - /* swap size statistics too, since new rel has freshly-updated stats */ { int32 swap_pages; @@ -2313,9 +2300,8 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, target_is_pg_class, swap_toast_by_content, is_internal, - frozenXid, - cutoffMulti, - mapped_tables); + mapped_tables, + NULL); } else { @@ -2416,9 +2402,8 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, target_is_pg_class, swap_toast_by_content, is_internal, - InvalidTransactionId, - InvalidMultiXactId, - mapped_tables); + mapped_tables, + NULL); } /* Clean up. */ @@ -2428,6 +2413,36 @@ swap_relation_files(Oid r1, Oid r2, bool target_is_pg_class, table_close(relRelation, RowExclusiveLock); } +/* + * Update relfrozenxid and relminmxid attributes of pg_class entry, specified + * by OID. + */ +static void +update_relation_cutoffs(Oid relid, TransactionId frozenXid, + MultiXactId cutoffMulti) +{ + Relation relRelation; + HeapTuple reltup; + Form_pg_class relform; + CatalogIndexState indstate; + + relRelation = table_open(RelationRelationId, RowExclusiveLock); + reltup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(relid)); + if (!HeapTupleIsValid(reltup)) + elog(ERROR, "cache lookup failed for relation %u", relid); + relform = (Form_pg_class) GETSTRUCT(reltup); + relform->relfrozenxid = frozenXid; + relform->relminmxid = cutoffMulti; + + indstate = CatalogOpenIndexes(relRelation); + CatalogTupleUpdateWithInfo(relRelation, &reltup->t_self, reltup, + indstate); + CatalogCloseIndexes(indstate); + + heap_freetuple(reltup); + table_close(relRelation, RowExclusiveLock); +} + /* * Remove the transient table that was built by make_new_heap, and finish * cleaning up (including rebuilding all indexes on the old heap). @@ -2445,6 +2460,7 @@ finish_heap_swap(Oid OIDOldHeap, Oid OIDNewHeap, { ObjectAddress object; Oid mapped_tables[4]; + Oid oid_old_toastid; int i; /* Report that we are now swapping relation files */ @@ -2461,7 +2477,7 @@ finish_heap_swap(Oid OIDOldHeap, Oid OIDNewHeap, swap_relation_files(OIDOldHeap, OIDNewHeap, (OIDOldHeap == RelationRelationId), swap_toast_by_content, is_internal, - frozenXid, cutoffMulti, mapped_tables); + mapped_tables, &oid_old_toastid); /* * If it's a system catalog, queue a sinval message to flush all catcaches @@ -2517,37 +2533,20 @@ finish_heap_swap(Oid OIDOldHeap, Oid OIDNewHeap, PROGRESS_REPACK_PHASE_FINAL_CLEANUP); /* - * If the relation being rebuilt is pg_class, swap_relation_files() - * couldn't update pg_class's own pg_class entry (check comments in - * swap_relation_files()), thus relfrozenxid was not updated. That's - * annoying because a potential reason for doing a VACUUM FULL is a - * imminent or actual anti-wraparound shutdown. So, now that we can - * access the new relation using its indices, update relfrozenxid. - * pg_class doesn't have a toast relation, so we don't need to update the - * corresponding toast relation. Not that there's little point moving all - * relfrozenxid updates here since swap_relation_files() needs to write to - * pg_class for non-mapped relations anyway. + * Update relfrozenxid and relminmxid. CCI is needed to see the most + * recent pg_class tuple, created by swap_relation_files(). However that + * CCI makes us use the new relation file. Therefore we cannot do this + * before indexes have been rebuilt too. */ - if (OIDOldHeap == RelationRelationId) - { - Relation relRelation; - HeapTuple reltup; - Form_pg_class relform; - - relRelation = table_open(RelationRelationId, RowExclusiveLock); - - reltup = SearchSysCacheCopy1(RELOID, ObjectIdGetDatum(OIDOldHeap)); - if (!HeapTupleIsValid(reltup)) - elog(ERROR, "cache lookup failed for relation %u", OIDOldHeap); - relform = (Form_pg_class) GETSTRUCT(reltup); - - relform->relfrozenxid = frozenXid; - relform->relminmxid = cutoffMulti; - - CatalogTupleUpdate(relRelation, &reltup->t_self, reltup); + CommandCounterIncrement(); + update_relation_cutoffs(OIDOldHeap, frozenXid, cutoffMulti); - table_close(relRelation, RowExclusiveLock); - } + /* + * The same for TOAST, if needed. In the swap-toast-links case, the new + * the toast table should already have relfrozenxid set to RecentXmin. + */ + if (OidIsValid(oid_old_toastid) && swap_toast_by_content) + update_relation_cutoffs(oid_old_toastid, frozenXid, cutoffMulti); /* Destroy new heap with old filenumber */ object.classId = RelationRelationId; @@ -4133,9 +4132,8 @@ rebuild_relation_finish_concurrent(Relation NewHeap, Relation OldHeap, (old_table_oid == RelationRelationId), false, /* swap_toast_by_content */ true, - InvalidTransactionId, - InvalidMultiXactId, - mapped_tables); + mapped_tables, + NULL); #ifdef USE_ASSERT_CHECKING -- 2.52.0 --=-=-= Content-Type: text/plain Content-Disposition: attachment; filename=v02-0008-Make-REPACK-CONCURRENTLY-MVCC-safe.patch ^ permalink raw reply [nested|flat] 268+ messages in thread
end of thread, other threads:[~2026-07-15 08:37 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 <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2025-07-26 00:57 [PATCH v12 2/3] Use time stamp counter to measure time on Linux/x86 Lukas Fittl <[email protected]> 2026-07-15 08:37 [PATCH 7/8] Decouple updating of freezing information from swap_relation_files(). Antonin Houska <[email protected]>
This inbox is served by agora; see mirroring instructions for how to clone and mirror all data and code used for this inbox