#include "hash.h"
#include "os/os.h"
-#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
+#if defined(ARCH_HAVE_CPU_CLOCK)
+#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
static unsigned long cycles_per_usec;
-static unsigned long inv_cycles_per_usec;
-static uint64_t max_cycles_for_mult;
-static unsigned long long cycles_start, cycles_wrap;
+static unsigned long long cycles_start;
+static unsigned long long clock_mult;
+static unsigned long long max_cycles_mask;
+static unsigned long long nsecs_for_max_cycles;
+static unsigned int clock_shift;
+static unsigned int max_cycles_shift;
+#define MAX_CLOCK_SEC 60*60
+#endif
+#ifdef ARCH_CPU_CLOCK_WRAPS
+static unsigned int cycles_wrap;
+#endif
#endif
int tsc_reliable = 0;
#ifdef CONFIG_CLOCK_GETTIME
static int fill_clock_gettime(struct timespec *ts)
{
-#ifdef CONFIG_CLOCK_MONOTONIC
+#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
+ return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
+#elif defined(CONFIG_CLOCK_MONOTONIC)
return clock_gettime(CLOCK_MONOTONIC, ts);
#else
return clock_gettime(CLOCK_REALTIME, ts);
}
#endif
-static void __fio_gettime(struct timeval *tp)
+static void __fio_gettime(struct timespec *tp)
{
switch (fio_clock_source) {
#ifdef CONFIG_GETTIMEOFDAY
- case CS_GTOD:
- gettimeofday(tp, NULL);
+ case CS_GTOD: {
+ struct timeval tv;
+ gettimeofday(&tv, NULL);
+
+ tp->tv_sec = tv.tv_sec;
+ tp->tv_nsec = tv.tv_usec * 1000;
break;
+ }
#endif
#ifdef CONFIG_CLOCK_GETTIME
case CS_CGETTIME: {
- struct timespec ts;
-
- if (fill_clock_gettime(&ts) < 0) {
+ if (fill_clock_gettime(tp) < 0) {
log_err("fio: clock_gettime fails\n");
assert(0);
}
-
- tp->tv_sec = ts.tv_sec;
- tp->tv_usec = ts.tv_nsec / 1000;
break;
}
#endif
#ifdef ARCH_HAVE_CPU_CLOCK
case CS_CPUCLOCK: {
- uint64_t usecs, t;
+ uint64_t nsecs, t, multiples;
struct tv_valid *tv;
#ifdef CONFIG_TLS_THREAD
#endif
t = get_cpu_clock();
+#ifdef ARCH_CPU_CLOCK_WRAPS
if (t < cycles_start && !cycles_wrap)
cycles_wrap = 1;
else if (cycles_wrap && t >= cycles_start && !tv->warned) {
log_err("fio: double CPU clock wrap\n");
tv->warned = 1;
}
-
- t -= cycles_start;
- tv->last_cycles = t;
- tv->last_tv_valid = 1;
+#endif
#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
- usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
+ nsecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC * 1000;
#else
- if (t < max_cycles_for_mult)
- usecs = (t * inv_cycles_per_usec) / 16777216UL;
- else
- usecs = t / cycles_per_usec;
+ t -= cycles_start;
+ multiples = t >> max_cycles_shift;
+ nsecs = multiples * nsecs_for_max_cycles;
+ nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift;
#endif
- tp->tv_sec = usecs / 1000000;
- tp->tv_usec = usecs % 1000000;
+ tv->last_cycles = t;
+ tv->last_tv_valid = 1;
+
+ tp->tv_sec = nsecs / 1000000000ULL;
+ tp->tv_nsec = nsecs % 1000000000ULL;
break;
}
#endif
}
#ifdef FIO_DEBUG_TIME
-void fio_gettime(struct timeval *tp, void *caller)
+void fio_gettime(struct timespec *tp, void *caller)
#else
-void fio_gettime(struct timeval *tp, void fio_unused *caller)
+void fio_gettime(struct timespec *tp, void fio_unused *caller)
#endif
{
#ifdef FIO_DEBUG_TIME
#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
static unsigned long get_cycles_per_usec(void)
{
- struct timeval s, e;
+ struct timespec s, e;
uint64_t c_s, c_e;
enum fio_cs old_cs = fio_clock_source;
+ uint64_t elapsed;
#ifdef CONFIG_CLOCK_GETTIME
fio_clock_source = CS_CGETTIME;
c_s = get_cpu_clock();
do {
- uint64_t elapsed;
-
__fio_gettime(&e);
elapsed = utime_since(&s, &e);
} while (1);
fio_clock_source = old_cs;
- return (c_e - c_s + 127) >> 7;
+ return (c_e - c_s) / elapsed;
}
#define NR_TIME_ITERS 50
{
double delta, mean, S;
uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
- int i, samples;
+ int i, samples, sft = 0;
+ unsigned long long tmp, max_ticks, max_mult;
cycles[0] = get_cycles_per_usec();
S = delta = mean = 0.0;
}
S /= (double) NR_TIME_ITERS;
- mean /= 10.0;
for (i = 0; i < NR_TIME_ITERS; i++)
- dprint(FD_TIME, "cycles[%d]=%llu\n", i,
- (unsigned long long) cycles[i] / 10);
+ dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
avg /= samples;
- avg = (avg + 5) / 10;
- minc /= 10;
- maxc /= 10;
+ cycles_per_usec = avg;
dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
(unsigned long long) minc,
(unsigned long long) maxc, mean, S);
- cycles_per_usec = avg;
- inv_cycles_per_usec = 16777216UL / cycles_per_usec;
- max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
- dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
+ max_ticks = MAX_CLOCK_SEC * cycles_per_usec * 1000000ULL;
+ max_mult = ULLONG_MAX / max_ticks;
+ dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, max_mult=%llu\n",
+ max_ticks, __builtin_clzll(max_ticks), max_mult);
+
+ /*
+ * Find the largest shift count that will produce
+ * a multiplier that does not exceed max_mult
+ */
+ tmp = max_mult * cycles_per_usec / 1000;
+ while (tmp > 1) {
+ tmp >>= 1;
+ sft++;
+ dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft);
+ }
+
+ clock_shift = sft;
+ clock_mult = (1ULL << sft) * 1000 / cycles_per_usec;
+ dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift, clock_mult);
+
+ // Find the greatest power of 2 clock ticks that is less than the ticks in MAX_CLOCK_SEC_2STAGE
+ max_cycles_shift = max_cycles_mask = 0;
+ tmp = MAX_CLOCK_SEC * 1000000ULL * cycles_per_usec;
+ dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
+ while (tmp > 1) {
+ tmp >>= 1;
+ max_cycles_shift++;
+ dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
+ }
+ // if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_usec here we will
+ // have a discontinuity every (1ULL << max_cycles_shift) cycles
+ nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult) >> clock_shift;
+
+ // Use a bitmask to calculate ticks % (1ULL << max_cycles_shift)
+ for (tmp = 0; tmp < max_cycles_shift; tmp++)
+ max_cycles_mask |= 1ULL << tmp;
+
+ dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, nsecs_for_max_cycles=%llu, max_cycles_mask=%016llx\n",
+ max_cycles_shift, (1ULL << max_cycles_shift),
+ nsecs_for_max_cycles, max_cycles_mask);
+
cycles_start = get_cpu_clock();
dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
return 0;
log_info("fio: clocksource=cpu may not be reliable\n");
}
-uint64_t utime_since(const struct timeval *s, const struct timeval *e)
+uint64_t ntime_since(const struct timespec *s, const struct timespec *e)
+{
+ int64_t sec, nsec;
+
+ sec = e->tv_sec - s->tv_sec;
+ nsec = e->tv_nsec - s->tv_nsec;
+ if (sec > 0 && nsec < 0) {
+ sec--;
+ nsec += 1000000000LL;
+ }
+
+ /*
+ * time warp bug on some kernels?
+ */
+ if (sec < 0 || (sec == 0 && nsec < 0))
+ return 0;
+
+ return nsec + (sec * 1000000000LL);
+}
+
+uint64_t utime_since(const struct timespec *s, const struct timespec *e)
{
- long sec, usec;
- uint64_t ret;
+ int64_t sec, usec;
sec = e->tv_sec - s->tv_sec;
- usec = e->tv_usec - s->tv_usec;
+ usec = (e->tv_nsec - s->tv_nsec) / 1000;
if (sec > 0 && usec < 0) {
sec--;
usec += 1000000;
if (sec < 0 || (sec == 0 && usec < 0))
return 0;
- ret = sec * 1000000ULL + usec;
-
- return ret;
+ return usec + (sec * 1000000);
}
-uint64_t utime_since_now(const struct timeval *s)
+uint64_t utime_since_now(const struct timespec *s)
{
- struct timeval t;
+ struct timespec t;
+#ifdef FIO_DEBUG_TIME
+ void *p = __builtin_return_address(0);
+ fio_gettime(&t, p);
+#else
fio_gettime(&t, NULL);
+#endif
+
return utime_since(s, &t);
}
-uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
+uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
{
- long sec, usec, ret;
+ int64_t sec, usec;
sec = e->tv_sec - s->tv_sec;
- usec = e->tv_usec - s->tv_usec;
+ usec = (e->tv_usec - s->tv_usec);
if (sec > 0 && usec < 0) {
sec--;
usec += 1000000;
if (sec < 0 || (sec == 0 && usec < 0))
return 0;
- sec *= 1000UL;
- usec /= 1000UL;
- ret = sec + usec;
-
- return ret;
+ sec *= 1000;
+ usec /= 1000;
+ return sec + usec;
}
-uint64_t mtime_since_now(const struct timeval *s)
+uint64_t mtime_since_now(const struct timespec *s)
{
- struct timeval t;
+ struct timespec t;
+#ifdef FIO_DEBUG_TIME
void *p = __builtin_return_address(0);
fio_gettime(&t, p);
+#else
+ fio_gettime(&t, NULL);
+#endif
+
return mtime_since(s, &t);
}
-uint64_t time_since_now(const struct timeval *s)
+uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
+{
+ int64_t sec, usec;
+
+ sec = e->tv_sec - s->tv_sec;
+ usec = (e->tv_nsec - s->tv_nsec) / 1000;
+ if (sec > 0 && usec < 0) {
+ sec--;
+ usec += 1000000;
+ }
+
+ if (sec < 0 || (sec == 0 && usec < 0))
+ return 0;
+
+ sec *= 1000;
+ usec /= 1000;
+ return sec + usec;
+}
+
+uint64_t time_since_now(const struct timespec *s)
{
return mtime_since_now(s) / 1000;
}
struct clock_entry *c;
os_cpu_mask_t cpu_mask;
uint32_t last_seq;
+ unsigned long long first;
int i;
- memset(&cpu_mask, 0, sizeof(cpu_mask));
+ if (fio_cpuset_init(&cpu_mask)) {
+ int __err = errno;
+
+ log_err("clock cpuset init failed: %s\n", strerror(__err));
+ goto err_out;
+ }
+
fio_cpu_set(&cpu_mask, t->cpu);
if (fio_setaffinity(gettid(), cpu_mask) == -1) {
int __err = errno;
log_err("clock setaffinity failed: %s\n", strerror(__err));
- return (void *) 1;
+ goto err;
}
pthread_mutex_lock(&t->lock);
pthread_mutex_unlock(&t->started);
+ first = get_cpu_clock();
last_seq = 0;
c = &t->entries[0];
for (i = 0; i < t->nr_entries; i++, c++) {
unsigned long long clocks;
clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
- log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu, clocks);
+ log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
+ clocks, first);
}
/*
* indefinitely. Check for that and return failure.
*/
if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
- return (void *) 1;
+ goto err;
+ fio_cpuset_exit(&cpu_mask);
return NULL;
+err:
+ fio_cpuset_exit(&cpu_mask);
+err_out:
+ return (void *) 1;
}
static int clock_cmp(const void *p1, const void *p2)
qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
+ /* silence silly gcc */
+ prev = NULL;
for (failed = i = 0; i < tentries; i++) {
this = &entries[i];
projects / fio.git / blobdiff