From 0399f5fc547ef035c4eb5e383f30b28ae73d936e Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Thu, 16 Jul 2015 11:27:00 +0200 Subject: erts: Refactor perf counter internal interface perf counter is now part of the function pointer interface and also the function returns the value instead of writing to a memory buffer. --- erts/emulator/sys/unix/erl_unix_sys.h | 35 +++++------ erts/emulator/sys/unix/sys_time.c | 112 ++++++++++++++++++++++++++++++++++ 2 files changed, 128 insertions(+), 19 deletions(-) (limited to 'erts/emulator/sys/unix') diff --git a/erts/emulator/sys/unix/erl_unix_sys.h b/erts/emulator/sys/unix/erl_unix_sys.h index 3a03d6be49..8b1822ca9f 100644 --- a/erts/emulator/sys/unix/erl_unix_sys.h +++ b/erts/emulator/sys/unix/erl_unix_sys.h @@ -161,6 +161,7 @@ typedef long long ErtsSysHrTime; #endif typedef ErtsMonotonicTime ErtsSystemTime; +typedef ErtsSysHrTime ErtsSysPerfCounter; #define ERTS_MONOTONIC_TIME_MIN (((ErtsMonotonicTime) 1) << 63) #define ERTS_MONOTONIC_TIME_MAX (~ERTS_MONOTONIC_TIME_MIN) @@ -209,6 +210,7 @@ ErtsSystemTime erts_os_system_time(void); * It may or may not be monotonic. */ ErtsSysHrTime erts_sys_hrtime(void); +#define ERTS_HRTIME_UNIT (1000*1000*1000) struct erts_sys_time_read_only_data__ { #ifdef ERTS_OS_MONOTONIC_INLINE_FUNC_PTR_CALL__ @@ -217,6 +219,8 @@ struct erts_sys_time_read_only_data__ { #ifdef ERTS_OS_TIMES_INLINE_FUNC_PTR_CALL__ void (*os_times)(ErtsMonotonicTime *, ErtsSystemTime *); #endif + ErtsSysPerfCounter (*perf_counter)(void); + ErtsSysPerfCounter perf_counter_unit; int ticks_per_sec; }; @@ -273,25 +277,18 @@ erts_os_times(ErtsMonotonicTime *mtimep, ErtsSystemTime *stimep) * Functions for getting the performance counter */ -#if defined(__x86_64__) - /* available on all x86_64. Best if used when we have constant_tsc and - nonstop_tsc cpu features. It may have been a good idea to put the - cpuid instruction before the rdtsc, but I decided against it - because it is not really needed for msacc, and it slows it down by - quite a bit. As a result though, this timestamp becomes much less - accurate as it might be re-ordered to be executed way before this - function is called. - */ -#define sys_perf_counter(ts) do { \ - __asm__ __volatile__ ("rdtsc\n\t" \ - "shl $32, %%rdx\n\t" \ - "or %%rdx, %0" : "=a" (*ts) : : "rdx"); \ - } while(0) -#define SYS_PERF_COUNTER_UNIT erts_perf_counter_unit() -#else -#define sys_perf_counter(ts) *(ts) = erts_sys_hrtime() -#define SYS_PERF_COUNTER_UNIT 1000000000LL -#endif +ERTS_GLB_INLINE ErtsSysPerfCounter erts_sys_perf_counter(void); +#define erts_sys_perf_counter_unit() erts_sys_time_data__.r.o.perf_counter_unit + +#if ERTS_GLB_INLINE_INCL_FUNC_DEF + +ERTS_GLB_INLINE ErtsSysPerfCounter +erts_sys_perf_counter() +{ + return (*erts_sys_time_data__.r.o.perf_counter)(); +} + +#endif /* ERTS_GLB_INLINE_INCL_FUNC_DEF */ /* * Functions for measuring CPU time diff --git a/erts/emulator/sys/unix/sys_time.c b/erts/emulator/sys/unix/sys_time.c index 6fc4fc7dc4..9738a8c352 100644 --- a/erts/emulator/sys/unix/sys_time.c +++ b/erts/emulator/sys/unix/sys_time.c @@ -65,6 +65,8 @@ # include #endif +static void init_perf_counter(void); + /******************* Routines for time measurement *********************/ #undef ERTS_SYS_TIME_INTERNAL_STATE_WRITE_FREQ__ @@ -404,6 +406,8 @@ sys_init_time(ErtsSysInitTimeResult *init_resp) # error Missing erts_os_system_time() implementation #endif + init_perf_counter(); + } void @@ -908,6 +912,114 @@ erts_os_times(ErtsMonotonicTime *mtimep, ErtsSystemTime *stimep) #endif +/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\ + * Performance counter functions * +\* */ + + +/* What resolution to spin to in micro seconds */ +#define RESOLUTION 100 +/* How many iterations to spin */ +#define ITERATIONS 1 +/* How many significant figures to round to */ +#define SIGFIGS 3 + +static ErtsSysPerfCounter calculate_perf_counter_unit(void) { + int i; + ErtsSysPerfCounter pre, post; + double value = 0; + double round_factor; +#if defined(HAVE_GETHRTIME) && defined(GETHRTIME_WITH_CLOCK_GETTIME) + struct timespec basetime,comparetime; +#define __GETTIME(arg) clock_gettime(CLOCK_MONOTONIC,arg) +#define __GETUSEC(arg) (arg.tv_nsec / 1000) +#else + SysTimeval basetime,comparetime; +#define __GETTIME(arg) sys_gettimeofday(arg) +#define __GETUSEC(arg) arg.tv_usec +#endif + + for (i = 0; i < ITERATIONS; i++) { + /* Make sure usec just flipped over at current resolution */ + __GETTIME(&basetime); + do { + __GETTIME(&comparetime); + } while ((__GETUSEC(basetime) / RESOLUTION) == (__GETUSEC(comparetime) / RESOLUTION)); + + pre = erts_sys_perf_counter(); + + __GETTIME(&basetime); + do { + __GETTIME(&comparetime); + } while ((__GETUSEC(basetime) / RESOLUTION) == (__GETUSEC(comparetime) / RESOLUTION)); + + post = erts_sys_perf_counter(); + + value += post - pre; + } + /* After this value is ticks per us */ + value /= (RESOLUTION*ITERATIONS); + + /* We round to 3 significant figures */ + round_factor = pow(10.0, SIGFIGS - ceil(log10(value))); + value = ((ErtsSysPerfCounter)(value * round_factor + 0.5)) / round_factor; + + /* convert to ticks per second */ + return 1000000 * value; +} + +static int have_rdtscp(void) +{ +#if defined(ETHR_X86_RUNTIME_CONF__) + /* On early x86 cpu's the tsc varies with + the current speed of the cpu, which means that the time per + tick vary depending on the current load of the cpu. We do not + want this as it would give very scewed numbers when the cpu is + mostly idle. + If we have the rdtscp feature it is a sign that the cpu is + relatively modern, and thus the tsc quite stable so we use it then. + + If this test is not good enough, I don't know what we'll do. + Maybe fallback on erts_sys_hrtime always, but that would be a shame as + rdtsc is about 3 times faster than hrtime... */ + return ETHR_X86_RUNTIME_CONF_HAVE_RDTSCP__; +#else + return 0; +#endif +} + +static ErtsSysPerfCounter rdtsc(void) +{ + /* It may have been a good idea to put the cpuid instruction before + the rdtsc, but I decided against it because it is not really + needed for msacc, and it slows it down by quite a bit (5-7 times slower). + As a result though, this timestamp becomes much less + accurate as it might be re-ordered to be executed way before or after this + function is called. + */ + ErtsSysPerfCounter ts; +#if defined(__x86_64__) + __asm__ __volatile__ ("rdtsc\n\t" + "shl $32, %%rdx\n\t" + "or %%rdx, %0" : "=a" (ts) : : "rdx"); +#elif defined(__i386__) + __asm__ __volatile__ ("rdtsc\n\t" + : "=A" (ts) ); +#endif + return ts; +} + +static void init_perf_counter(void) +{ + if (have_rdtscp()) { + erts_sys_time_data__.r.o.perf_counter = rdtsc; + erts_sys_time_data__.r.o.perf_counter_unit = calculate_perf_counter_unit(); + } else { + erts_sys_time_data__.r.o.perf_counter = erts_sys_hrtime; + erts_sys_time_data__.r.o.perf_counter_unit = ERTS_HRTIME_UNIT; + } +} + /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ #ifdef HAVE_GETHRVTIME_PROCFS_IOCTL -- cgit v1.2.3