16 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
17 static unsigned long cycles_per_usec;
18 static unsigned long inv_cycles_per_usec;
27 #ifdef CONFIG_TLS_THREAD
28 static __thread struct tv_valid static_tv_valid;
30 static pthread_key_t tv_tls_key;
33 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
34 int fio_clock_source_set = 0;
35 static enum fio_cs fio_clock_source_inited = CS_INVAL;
40 #define HASH_SIZE (1 << HASH_BITS)
42 static struct flist_head hash[HASH_SIZE];
43 static int gtod_inited;
46 struct flist_head list;
51 static struct gtod_log *find_hash(void *caller)
53 unsigned long h = hash_ptr(caller, HASH_BITS);
54 struct flist_head *entry;
56 flist_for_each(entry, &hash[h]) {
57 struct gtod_log *log = flist_entry(entry, struct gtod_log,
60 if (log->caller == caller)
67 static struct gtod_log *find_log(void *caller)
69 struct gtod_log *log = find_hash(caller);
74 log = malloc(sizeof(*log));
75 INIT_FLIST_HEAD(&log->list);
79 h = hash_ptr(caller, HASH_BITS);
80 flist_add_tail(&log->list, &hash[h]);
86 static void gtod_log_caller(void *caller)
89 struct gtod_log *log = find_log(caller);
95 static void fio_exit fio_dump_gtod(void)
97 unsigned long total_calls = 0;
100 for (i = 0; i < HASH_SIZE; i++) {
101 struct flist_head *entry;
102 struct gtod_log *log;
104 flist_for_each(entry, &hash[i]) {
105 log = flist_entry(entry, struct gtod_log, list);
107 printf("function %p, calls %lu\n", log->caller,
109 total_calls += log->calls;
113 printf("Total %lu gettimeofday\n", total_calls);
116 static void fio_init gtod_init(void)
120 for (i = 0; i < HASH_SIZE; i++)
121 INIT_FLIST_HEAD(&hash[i]);
126 #endif /* FIO_DEBUG_TIME */
128 #ifdef CONFIG_CLOCK_GETTIME
129 static int fill_clock_gettime(struct timespec *ts)
131 #ifdef CONFIG_CLOCK_MONOTONIC
132 return clock_gettime(CLOCK_MONOTONIC, ts);
134 return clock_gettime(CLOCK_REALTIME, ts);
139 static void __fio_gettime(struct timeval *tp)
143 #ifdef CONFIG_TLS_THREAD
144 tv = &static_tv_valid;
146 tv = pthread_getspecific(tv_tls_key);
149 switch (fio_clock_source) {
150 #ifdef CONFIG_GETTIMEOFDAY
152 gettimeofday(tp, NULL);
155 #ifdef CONFIG_CLOCK_GETTIME
159 if (fill_clock_gettime(&ts) < 0) {
160 log_err("fio: clock_gettime fails\n");
164 tp->tv_sec = ts.tv_sec;
165 tp->tv_usec = ts.tv_nsec / 1000;
169 #ifdef ARCH_HAVE_CPU_CLOCK
174 if (t < tv->last_cycles && tv->last_tv_valid &&
176 log_err("fio: CPU clock going back in time\n");
181 tv->last_tv_valid = 1;
182 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
183 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
185 usecs = (t * inv_cycles_per_usec) / 16777216UL;
187 tp->tv_sec = usecs / 1000000;
188 tp->tv_usec = usecs % 1000000;
193 log_err("fio: invalid clock source %d\n", fio_clock_source);
198 #ifdef FIO_DEBUG_TIME
199 void fio_gettime(struct timeval *tp, void *caller)
201 void fio_gettime(struct timeval *tp, void fio_unused *caller)
204 #ifdef FIO_DEBUG_TIME
206 caller = __builtin_return_address(0);
208 gtod_log_caller(caller);
210 if (fio_unlikely(fio_tv)) {
211 memcpy(tp, fio_tv, sizeof(*tp));
218 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
219 static unsigned long get_cycles_per_usec(void)
223 enum fio_cs old_cs = fio_clock_source;
225 #ifdef CONFIG_CLOCK_GETTIME
226 fio_clock_source = CS_CGETTIME;
228 fio_clock_source = CS_GTOD;
232 c_s = get_cpu_clock();
238 elapsed = utime_since(&s, &e);
239 if (elapsed >= 1280) {
240 c_e = get_cpu_clock();
245 fio_clock_source = old_cs;
246 return (c_e - c_s + 127) >> 7;
249 #define NR_TIME_ITERS 50
251 static int calibrate_cpu_clock(void)
253 double delta, mean, S;
254 uint64_t avg, cycles[NR_TIME_ITERS];
257 cycles[0] = get_cycles_per_usec();
258 S = delta = mean = 0.0;
259 for (i = 0; i < NR_TIME_ITERS; i++) {
260 cycles[i] = get_cycles_per_usec();
261 delta = cycles[i] - mean;
263 mean += delta / (i + 1.0);
264 S += delta * (cycles[i] - mean);
269 * The most common platform clock breakage is returning zero
270 * indefinitely. Check for that and return failure.
272 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
275 S = sqrt(S / (NR_TIME_ITERS - 1.0));
278 for (i = 0; i < NR_TIME_ITERS; i++) {
279 double this = cycles[i];
281 if ((fmax(this, mean) - fmin(this, mean)) > S)
287 S /= (double) NR_TIME_ITERS;
290 for (i = 0; i < NR_TIME_ITERS; i++)
291 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
292 (unsigned long long) cycles[i] / 10);
295 avg = (avg + 5) / 10;
296 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
297 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
299 cycles_per_usec = avg;
300 inv_cycles_per_usec = 16777216UL / cycles_per_usec;
301 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
305 static int calibrate_cpu_clock(void)
307 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
313 #endif // ARCH_HAVE_CPU_CLOCK
315 #ifndef CONFIG_TLS_THREAD
316 void fio_local_clock_init(int is_thread)
320 t = calloc(1, sizeof(*t));
321 if (pthread_setspecific(tv_tls_key, t)) {
322 log_err("fio: can't set TLS key\n");
327 static void kill_tv_tls_key(void *data)
332 void fio_local_clock_init(int is_thread)
337 void fio_clock_init(void)
339 if (fio_clock_source == fio_clock_source_inited)
342 #ifndef CONFIG_TLS_THREAD
343 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
344 log_err("fio: can't create TLS key\n");
347 fio_clock_source_inited = fio_clock_source;
349 if (calibrate_cpu_clock())
353 * If the arch sets tsc_reliable != 0, then it must be good enough
354 * to use as THE clock source. For x86 CPUs, this means the TSC
355 * runs at a constant rate and is synced across CPU cores.
358 if (!fio_clock_source_set)
359 fio_clock_source = CS_CPUCLOCK;
360 } else if (fio_clock_source == CS_CPUCLOCK)
361 log_info("fio: clocksource=cpu may not be reliable\n");
364 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
369 sec = e->tv_sec - s->tv_sec;
370 usec = e->tv_usec - s->tv_usec;
371 if (sec > 0 && usec < 0) {
377 * time warp bug on some kernels?
379 if (sec < 0 || (sec == 0 && usec < 0))
382 ret = sec * 1000000ULL + usec;
387 uint64_t utime_since_now(const struct timeval *s)
391 fio_gettime(&t, NULL);
392 return utime_since(s, &t);
395 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
399 sec = e->tv_sec - s->tv_sec;
400 usec = e->tv_usec - s->tv_usec;
401 if (sec > 0 && usec < 0) {
406 if (sec < 0 || (sec == 0 && usec < 0))
416 uint64_t mtime_since_now(const struct timeval *s)
419 void *p = __builtin_return_address(0);
422 return mtime_since(s, &t);
425 uint64_t time_since_now(const struct timeval *s)
427 return mtime_since_now(s) / 1000;
430 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
433 #define CLOCK_ENTRIES 100000
441 struct clock_thread {
444 pthread_mutex_t lock;
445 pthread_mutex_t started;
447 struct clock_entry *entries;
450 static inline uint32_t atomic32_inc_return(uint32_t *seq)
452 return 1 + __sync_fetch_and_add(seq, 1);
455 static void *clock_thread_fn(void *data)
457 struct clock_thread *t = data;
458 struct clock_entry *c;
459 os_cpu_mask_t cpu_mask;
463 memset(&cpu_mask, 0, sizeof(cpu_mask));
464 fio_cpu_set(&cpu_mask, t->cpu);
466 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
467 log_err("clock setaffinity failed\n");
471 pthread_mutex_lock(&t->lock);
472 pthread_mutex_unlock(&t->started);
476 for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
482 seq = atomic32_inc_return(t->seq);
485 tsc = get_cpu_clock();
486 } while (seq != *t->seq);
492 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
493 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
496 * The most common platform clock breakage is returning zero
497 * indefinitely. Check for that and return failure.
499 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
505 static int clock_cmp(const void *p1, const void *p2)
507 const struct clock_entry *c1 = p1;
508 const struct clock_entry *c2 = p2;
510 if (c1->seq == c2->seq)
511 log_err("cs: bug in atomic sequence!\n");
513 return c1->seq - c2->seq;
516 int fio_monotonic_clocktest(void)
518 struct clock_thread *cthreads;
519 unsigned int nr_cpus = cpus_online();
520 struct clock_entry *entries;
521 unsigned long tentries, failed = 0;
522 struct clock_entry *prev, *this;
526 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
529 fio_debug |= 1U << FD_TIME;
531 calibrate_cpu_clock();
533 fio_debug &= ~(1U << FD_TIME);
536 cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
537 tentries = CLOCK_ENTRIES * nr_cpus;
538 entries = malloc(tentries * sizeof(struct clock_entry));
540 log_info("cs: Testing %u CPUs\n", nr_cpus);
542 for (i = 0; i < nr_cpus; i++) {
543 struct clock_thread *t = &cthreads[i];
547 t->entries = &entries[i * CLOCK_ENTRIES];
548 pthread_mutex_init(&t->lock, NULL);
549 pthread_mutex_init(&t->started, NULL);
550 pthread_mutex_lock(&t->lock);
551 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
558 for (i = 0; i < nr_cpus; i++) {
559 struct clock_thread *t = &cthreads[i];
561 pthread_mutex_lock(&t->started);
564 for (i = 0; i < nr_cpus; i++) {
565 struct clock_thread *t = &cthreads[i];
567 pthread_mutex_unlock(&t->lock);
570 for (i = 0; i < nr_cpus; i++) {
571 struct clock_thread *t = &cthreads[i];
574 pthread_join(t->thread, &ret);
581 log_err("Clocksource test: %lu threads failed\n", failed);
585 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
587 for (failed = i = 0; i < tentries; i++) {
595 if (prev->tsc > this->tsc) {
596 uint64_t diff = prev->tsc - this->tsc;
598 log_info("cs: CPU clock mismatch (diff=%llu):\n",
599 (unsigned long long) diff);
600 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
601 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
609 log_info("cs: Failed: %lu\n", failed);
611 log_info("cs: Pass!\n");
618 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
620 int fio_monotonic_clocktest(void)
622 log_info("cs: current platform does not support CPU clocks\n");