diff options
Diffstat (limited to 'erts/emulator/sys/unix/sys_time.c')
-rw-r--r-- | erts/emulator/sys/unix/sys_time.c | 138 |
1 files changed, 126 insertions, 12 deletions
diff --git a/erts/emulator/sys/unix/sys_time.c b/erts/emulator/sys/unix/sys_time.c index 2e1914f564..8ce02506ab 100644 --- a/erts/emulator/sys/unix/sys_time.c +++ b/erts/emulator/sys/unix/sys_time.c @@ -65,6 +65,8 @@ # include <fcntl.h> #endif +static void init_perf_counter(void); + /******************* Routines for time measurement *********************/ #undef ERTS_SYS_TIME_INTERNAL_STATE_WRITE_FREQ__ @@ -342,7 +344,7 @@ sys_init_time(ErtsSysInitTimeResult *init_resp) * times() (CLK_TCK), the resolution is always one millisecond.. */ if ((erts_sys_time_data__.r.o.ticks_per_sec = TICKS_PER_SEC()) < 0) - erl_exit(ERTS_ABORT_EXIT, "Can't get clock ticks/sec\n"); + erts_exit(ERTS_ABORT_EXIT, "Can't get clock ticks/sec\n"); #if defined(OS_MONOTONIC_TIME_USING_TIMES) #if ERTS_COMPILE_TIME_MONOTONIC_TIME_UNIT @@ -404,6 +406,8 @@ sys_init_time(ErtsSysInitTimeResult *init_resp) # error Missing erts_os_system_time() implementation #endif + init_perf_counter(); + } void @@ -450,7 +454,7 @@ posix_clock_gettime(clockid_t id, char *name) if (clock_gettime(id, &ts) != 0) { int err = errno; char *errstr = err ? strerror(err) : "unknown"; - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_gettime(%s, _) failed: %s (%d)\n", name, errstr, err); } @@ -495,13 +499,13 @@ posix_clock_gettime_times(clockid_t mid, char *mname, if (mres != 0) { char *errstr = merr ? strerror(merr) : "unknown"; - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_gettime(%s, _) failed: %s (%d)\n", mname, errstr, merr); } if (sres != 0) { char *errstr = serr ? strerror(serr) : "unknown"; - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_gettime(%s, _) failed: %s (%d)\n", sname, errstr, serr); } @@ -674,7 +678,7 @@ mach_clocks_init(void) clck_srv_p = &internal_state.r.o.mach.clock.monotonic.srv; kret = host_get_clock_service(host, id, clck_srv_p); if (kret != KERN_SUCCESS) { - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "host_get_clock_service(_, %s, _) failed\n", name); } @@ -686,7 +690,7 @@ mach_clocks_init(void) clck_srv_p = &internal_state.r.o.mach.clock.wall.srv; kret = host_get_clock_service(host, id, clck_srv_p); if (kret != KERN_SUCCESS) { - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "host_get_clock_service(_, %s, _) failed\n", name); } @@ -695,7 +699,7 @@ mach_clocks_init(void) if (atexit(mach_clocks_fini) != 0) { int err = errno; char *errstr = err ? strerror(err) : "unknown"; - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "Failed to register mach_clocks_fini() " "for call at exit: %s (%d)\n", errstr, err); @@ -717,7 +721,7 @@ mach_clock_getres(ErtsMachClock *clk) (clock_attr_t) attr, &cnt); if (kret != KERN_SUCCESS || cnt != 1) { - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_get_attributes(%s, _) failed\n", clk->name); } @@ -735,7 +739,7 @@ mach_clock_get_time(ErtsMachClock *clk) kret = clock_get_time(clk->srv, &time_spec); if (kret != KERN_SUCCESS) - erl_exit(ERTS_ABORT_EXIT, "clock_get_time(%s, _) failed\n", clk->name); + erts_exit(ERTS_ABORT_EXIT, "clock_get_time(%s, _) failed\n", clk->name); return ERTS_TimeSpec2Sint64(&time_spec); } @@ -781,11 +785,11 @@ erts_os_times(ErtsMonotonicTime *mtimep, ErtsSystemTime *stimep) &sys_time_spec); if (mkret != KERN_SUCCESS) - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_get_time(%s, _) failed\n", internal_state.r.o.mach.clock.monotonic.name); if (skret != KERN_SUCCESS) - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "clock_get_time(%s, _) failed\n", internal_state.r.o.mach.clock.wall.name); @@ -850,7 +854,7 @@ erts_os_system_time(void) if (gettimeofday(&tv, NULL) != 0) { int err = errno; char *errstr = err ? strerror(err) : "unknown"; - erl_exit(ERTS_ABORT_EXIT, + erts_exit(ERTS_ABORT_EXIT, "gettimeofday(_, NULL) failed: %s (%d)\n", errstr, err); } @@ -908,10 +912,120 @@ 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. + The linux kernel seems to think that checking for constant and + reliable is enough to trust the counter so we do the same. + + 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_CONSTANT_TSC__ && + ETHR_X86_RUNTIME_CONF_HAVE_TSC_RELIABLE__; +#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 +/* The code below only has effect on solaris < 10, + needed in order for gehhrvtime to work properly */ int sys_start_hrvtime(void) { long msacct = PR_MSACCT; |