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;
23 struct timeval last_tv;
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 (tv && t < tv->last_cycles) {
175 dprint(FD_TIME, "CPU clock going back in time\n");
180 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
181 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
183 usecs = (t * inv_cycles_per_usec) / 16777216UL;
185 tp->tv_sec = usecs / 1000000;
186 tp->tv_usec = usecs % 1000000;
191 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)
206 #ifdef FIO_DEBUG_TIME
208 caller = __builtin_return_address(0);
210 gtod_log_caller(caller);
212 if (fio_unlikely(fio_tv)) {
213 memcpy(tp, fio_tv, sizeof(*tp));
217 tv = __fio_gettime(tp);
220 * If Linux is using the tsc clock on non-synced processors,
221 * sometimes time can appear to drift backwards. Fix that up.
224 if (tv->last_tv_valid) {
225 if (tp->tv_sec < tv->last_tv.tv_sec)
226 tp->tv_sec = tv->last_tv.tv_sec;
227 else if (tv->last_tv.tv_sec == tp->tv_sec &&
228 tp->tv_usec < tv->last_tv.tv_usec)
229 tp->tv_usec = tv->last_tv.tv_usec;
231 tv->last_tv_valid = 1;
232 memcpy(&tv->last_tv, tp, sizeof(*tp));
236 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
237 static unsigned long get_cycles_per_usec(void)
241 enum fio_cs old_cs = fio_clock_source;
243 #ifdef CONFIG_CLOCK_GETTIME
244 fio_clock_source = CS_CGETTIME;
246 fio_clock_source = CS_GTOD;
250 c_s = get_cpu_clock();
256 elapsed = utime_since(&s, &e);
257 if (elapsed >= 1280) {
258 c_e = get_cpu_clock();
263 fio_clock_source = old_cs;
264 return (c_e - c_s + 127) >> 7;
267 #define NR_TIME_ITERS 50
269 static int calibrate_cpu_clock(void)
271 double delta, mean, S;
272 uint64_t avg, cycles[NR_TIME_ITERS];
275 cycles[0] = get_cycles_per_usec();
276 S = delta = mean = 0.0;
277 for (i = 0; i < NR_TIME_ITERS; i++) {
278 cycles[i] = get_cycles_per_usec();
279 delta = cycles[i] - mean;
281 mean += delta / (i + 1.0);
282 S += delta * (cycles[i] - mean);
287 * The most common platform clock breakage is returning zero
288 * indefinitely. Check for that and return failure.
290 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
293 S = sqrt(S / (NR_TIME_ITERS - 1.0));
296 for (i = 0; i < NR_TIME_ITERS; i++) {
297 double this = cycles[i];
299 if ((fmax(this, mean) - fmin(this, mean)) > S)
305 S /= (double) NR_TIME_ITERS;
308 for (i = 0; i < NR_TIME_ITERS; i++)
309 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
310 (unsigned long long) cycles[i] / 10);
313 avg = (avg + 5) / 10;
314 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
315 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
317 cycles_per_usec = avg;
318 inv_cycles_per_usec = 16777216UL / cycles_per_usec;
319 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
323 static int calibrate_cpu_clock(void)
325 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
331 #endif // ARCH_HAVE_CPU_CLOCK
333 #ifndef CONFIG_TLS_THREAD
334 void fio_local_clock_init(int is_thread)
338 t = calloc(1, sizeof(*t));
339 if (pthread_setspecific(tv_tls_key, t))
340 log_err("fio: can't set TLS key\n");
343 static void kill_tv_tls_key(void *data)
348 void fio_local_clock_init(int is_thread)
353 void fio_clock_init(void)
355 if (fio_clock_source == fio_clock_source_inited)
358 #ifndef CONFIG_TLS_THREAD
359 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
360 log_err("fio: can't create TLS key\n");
363 fio_clock_source_inited = fio_clock_source;
365 if (calibrate_cpu_clock())
369 * If the arch sets tsc_reliable != 0, then it must be good enough
370 * to use as THE clock source. For x86 CPUs, this means the TSC
371 * runs at a constant rate and is synced across CPU cores.
374 if (!fio_clock_source_set)
375 fio_clock_source = CS_CPUCLOCK;
376 } else if (fio_clock_source == CS_CPUCLOCK)
377 log_info("fio: clocksource=cpu may not be reliable\n");
380 uint64_t utime_since(struct timeval *s, struct timeval *e)
385 sec = e->tv_sec - s->tv_sec;
386 usec = e->tv_usec - s->tv_usec;
387 if (sec > 0 && usec < 0) {
393 * time warp bug on some kernels?
395 if (sec < 0 || (sec == 0 && usec < 0))
398 ret = sec * 1000000ULL + usec;
403 uint64_t utime_since_now(struct timeval *s)
407 fio_gettime(&t, NULL);
408 return utime_since(s, &t);
411 uint64_t mtime_since(struct timeval *s, struct timeval *e)
415 sec = e->tv_sec - s->tv_sec;
416 usec = e->tv_usec - s->tv_usec;
417 if (sec > 0 && usec < 0) {
422 if (sec < 0 || (sec == 0 && usec < 0))
432 uint64_t mtime_since_now(struct timeval *s)
435 void *p = __builtin_return_address(0);
438 return mtime_since(s, &t);
441 uint64_t time_since_now(struct timeval *s)
443 return mtime_since_now(s) / 1000;
446 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
449 #define CLOCK_ENTRIES 100000
457 struct clock_thread {
460 pthread_mutex_t lock;
461 pthread_mutex_t started;
463 struct clock_entry *entries;
466 static inline uint32_t atomic32_inc_return(uint32_t *seq)
468 return 1 + __sync_fetch_and_add(seq, 1);
471 static void *clock_thread_fn(void *data)
473 struct clock_thread *t = data;
474 struct clock_entry *c;
475 os_cpu_mask_t cpu_mask;
479 memset(&cpu_mask, 0, sizeof(cpu_mask));
480 fio_cpu_set(&cpu_mask, t->cpu);
482 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
483 log_err("clock setaffinity failed\n");
487 pthread_mutex_lock(&t->lock);
488 pthread_mutex_unlock(&t->started);
492 for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
498 seq = atomic32_inc_return(t->seq);
501 tsc = get_cpu_clock();
502 } while (seq != *t->seq);
508 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
509 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
512 * The most common platform clock breakage is returning zero
513 * indefinitely. Check for that and return failure.
515 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
521 static int clock_cmp(const void *p1, const void *p2)
523 const struct clock_entry *c1 = p1;
524 const struct clock_entry *c2 = p2;
526 if (c1->seq == c2->seq)
527 log_err("cs: bug in atomic sequence!\n");
529 return c1->seq - c2->seq;
532 int fio_monotonic_clocktest(void)
534 struct clock_thread *threads;
535 unsigned int nr_cpus = cpus_online();
536 struct clock_entry *entries;
537 unsigned long tentries, failed;
538 struct clock_entry *prev, *this;
542 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
544 fio_debug |= 1U << FD_TIME;
545 calibrate_cpu_clock();
546 fio_debug &= ~(1U << FD_TIME);
548 threads = malloc(nr_cpus * sizeof(struct clock_thread));
549 tentries = CLOCK_ENTRIES * nr_cpus;
550 entries = malloc(tentries * sizeof(struct clock_entry));
552 log_info("cs: Testing %u CPUs\n", nr_cpus);
554 for (i = 0; i < nr_cpus; i++) {
555 struct clock_thread *t = &threads[i];
559 t->entries = &entries[i * CLOCK_ENTRIES];
560 pthread_mutex_init(&t->lock, NULL);
561 pthread_mutex_init(&t->started, NULL);
562 pthread_mutex_lock(&t->lock);
563 pthread_create(&t->thread, NULL, clock_thread_fn, t);
566 for (i = 0; i < nr_cpus; i++) {
567 struct clock_thread *t = &threads[i];
569 pthread_mutex_lock(&t->started);
572 for (i = 0; i < nr_cpus; i++) {
573 struct clock_thread *t = &threads[i];
575 pthread_mutex_unlock(&t->lock);
578 for (failed = i = 0; i < nr_cpus; i++) {
579 struct clock_thread *t = &threads[i];
582 pthread_join(t->thread, &ret);
589 log_err("Clocksource test: %lu threads failed\n", failed);
593 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
595 for (failed = i = 0; i < tentries; i++) {
603 if (prev->tsc > this->tsc) {
604 uint64_t diff = prev->tsc - this->tsc;
606 log_info("cs: CPU clock mismatch (diff=%llu):\n",
607 (unsigned long long) diff);
608 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
609 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
617 log_info("cs: Failed: %lu\n", failed);
619 log_info("cs: Pass!\n");
626 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
628 int fio_monotonic_clocktest(void)
630 log_info("cs: current platform does not support CPU clocks\n");