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;
19 static uint64_t max_cycles_for_mult;
20 static unsigned long long cycles_start, cycles_wrap;
29 #ifdef CONFIG_TLS_THREAD
30 static __thread struct tv_valid static_tv_valid;
32 static pthread_key_t tv_tls_key;
35 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
36 int fio_clock_source_set = 0;
37 static enum fio_cs fio_clock_source_inited = CS_INVAL;
42 #define HASH_SIZE (1 << HASH_BITS)
44 static struct flist_head hash[HASH_SIZE];
45 static int gtod_inited;
48 struct flist_head list;
53 static struct gtod_log *find_hash(void *caller)
55 unsigned long h = hash_ptr(caller, HASH_BITS);
56 struct flist_head *entry;
58 flist_for_each(entry, &hash[h]) {
59 struct gtod_log *log = flist_entry(entry, struct gtod_log,
62 if (log->caller == caller)
69 static void inc_caller(void *caller)
71 struct gtod_log *log = find_hash(caller);
76 log = malloc(sizeof(*log));
77 INIT_FLIST_HEAD(&log->list);
81 h = hash_ptr(caller, HASH_BITS);
82 flist_add_tail(&log->list, &hash[h]);
88 static void gtod_log_caller(void *caller)
94 static void fio_exit fio_dump_gtod(void)
96 unsigned long total_calls = 0;
99 for (i = 0; i < HASH_SIZE; i++) {
100 struct flist_head *entry;
101 struct gtod_log *log;
103 flist_for_each(entry, &hash[i]) {
104 log = flist_entry(entry, struct gtod_log, list);
106 printf("function %p, calls %lu\n", log->caller,
108 total_calls += log->calls;
112 printf("Total %lu gettimeofday\n", total_calls);
115 static void fio_init gtod_init(void)
119 for (i = 0; i < HASH_SIZE; i++)
120 INIT_FLIST_HEAD(&hash[i]);
125 #endif /* FIO_DEBUG_TIME */
127 #ifdef CONFIG_CLOCK_GETTIME
128 static int fill_clock_gettime(struct timespec *ts)
130 #ifdef CONFIG_CLOCK_MONOTONIC
131 return clock_gettime(CLOCK_MONOTONIC, ts);
133 return clock_gettime(CLOCK_REALTIME, ts);
138 static void __fio_gettime(struct timeval *tp)
142 #ifdef CONFIG_TLS_THREAD
143 tv = &static_tv_valid;
145 tv = pthread_getspecific(tv_tls_key);
148 switch (fio_clock_source) {
149 #ifdef CONFIG_GETTIMEOFDAY
151 gettimeofday(tp, NULL);
154 #ifdef CONFIG_CLOCK_GETTIME
158 if (fill_clock_gettime(&ts) < 0) {
159 log_err("fio: clock_gettime fails\n");
163 tp->tv_sec = ts.tv_sec;
164 tp->tv_usec = ts.tv_nsec / 1000;
168 #ifdef ARCH_HAVE_CPU_CLOCK
173 if (t < cycles_start && !cycles_wrap)
175 else if (cycles_wrap && t >= cycles_start) {
177 log_err("fio: double CPU clock wrap\n");
184 tv->last_tv_valid = 1;
185 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
186 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
188 if (t < max_cycles_for_mult)
189 usecs = (t * inv_cycles_per_usec) / 16777216UL;
191 usecs = t / cycles_per_usec;
193 tp->tv_sec = usecs / 1000000;
194 tp->tv_usec = usecs % 1000000;
199 log_err("fio: invalid clock source %d\n", fio_clock_source);
204 #ifdef FIO_DEBUG_TIME
205 void fio_gettime(struct timeval *tp, void *caller)
207 void fio_gettime(struct timeval *tp, void fio_unused *caller)
210 #ifdef FIO_DEBUG_TIME
212 caller = __builtin_return_address(0);
214 gtod_log_caller(caller);
216 if (fio_unlikely(fio_gettime_offload(tp)))
222 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
223 static unsigned long get_cycles_per_usec(void)
227 enum fio_cs old_cs = fio_clock_source;
229 #ifdef CONFIG_CLOCK_GETTIME
230 fio_clock_source = CS_CGETTIME;
232 fio_clock_source = CS_GTOD;
236 c_s = get_cpu_clock();
242 elapsed = utime_since(&s, &e);
243 if (elapsed >= 1280) {
244 c_e = get_cpu_clock();
249 fio_clock_source = old_cs;
250 return (c_e - c_s + 127) >> 7;
253 #define NR_TIME_ITERS 50
255 static int calibrate_cpu_clock(void)
257 double delta, mean, S;
258 uint64_t avg, cycles[NR_TIME_ITERS];
261 cycles[0] = get_cycles_per_usec();
262 S = delta = mean = 0.0;
263 for (i = 0; i < NR_TIME_ITERS; i++) {
264 cycles[i] = get_cycles_per_usec();
265 delta = cycles[i] - mean;
267 mean += delta / (i + 1.0);
268 S += delta * (cycles[i] - mean);
273 * The most common platform clock breakage is returning zero
274 * indefinitely. Check for that and return failure.
276 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
279 S = sqrt(S / (NR_TIME_ITERS - 1.0));
282 for (i = 0; i < NR_TIME_ITERS; i++) {
283 double this = cycles[i];
285 if ((fmax(this, mean) - fmin(this, mean)) > S)
291 S /= (double) NR_TIME_ITERS;
294 for (i = 0; i < NR_TIME_ITERS; i++)
295 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
296 (unsigned long long) cycles[i] / 10);
299 avg = (avg + 5) / 10;
300 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
301 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
303 cycles_per_usec = avg;
304 inv_cycles_per_usec = 16777216UL / cycles_per_usec;
305 max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
306 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
307 cycles_start = get_cpu_clock();
308 dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
312 static int calibrate_cpu_clock(void)
314 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
320 #endif // ARCH_HAVE_CPU_CLOCK
322 #ifndef CONFIG_TLS_THREAD
323 void fio_local_clock_init(int is_thread)
327 t = calloc(1, sizeof(*t));
328 if (pthread_setspecific(tv_tls_key, t)) {
329 log_err("fio: can't set TLS key\n");
334 static void kill_tv_tls_key(void *data)
339 void fio_local_clock_init(int is_thread)
344 void fio_clock_init(void)
346 if (fio_clock_source == fio_clock_source_inited)
349 #ifndef CONFIG_TLS_THREAD
350 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
351 log_err("fio: can't create TLS key\n");
354 fio_clock_source_inited = fio_clock_source;
356 if (calibrate_cpu_clock())
360 * If the arch sets tsc_reliable != 0, then it must be good enough
361 * to use as THE clock source. For x86 CPUs, this means the TSC
362 * runs at a constant rate and is synced across CPU cores.
365 if (!fio_clock_source_set)
366 fio_clock_source = CS_CPUCLOCK;
367 } else if (fio_clock_source == CS_CPUCLOCK)
368 log_info("fio: clocksource=cpu may not be reliable\n");
371 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
376 sec = e->tv_sec - s->tv_sec;
377 usec = e->tv_usec - s->tv_usec;
378 if (sec > 0 && usec < 0) {
384 * time warp bug on some kernels?
386 if (sec < 0 || (sec == 0 && usec < 0))
389 ret = sec * 1000000ULL + usec;
394 uint64_t utime_since_now(const struct timeval *s)
398 fio_gettime(&t, NULL);
399 return utime_since(s, &t);
402 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
406 sec = e->tv_sec - s->tv_sec;
407 usec = e->tv_usec - s->tv_usec;
408 if (sec > 0 && usec < 0) {
413 if (sec < 0 || (sec == 0 && usec < 0))
423 uint64_t mtime_since_now(const struct timeval *s)
426 void *p = __builtin_return_address(0);
429 return mtime_since(s, &t);
432 uint64_t time_since_now(const struct timeval *s)
434 return mtime_since_now(s) / 1000;
437 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
440 #define CLOCK_ENTRIES 100000
448 struct clock_thread {
451 pthread_mutex_t lock;
452 pthread_mutex_t started;
454 struct clock_entry *entries;
457 static inline uint32_t atomic32_inc_return(uint32_t *seq)
459 return 1 + __sync_fetch_and_add(seq, 1);
462 static void *clock_thread_fn(void *data)
464 struct clock_thread *t = data;
465 struct clock_entry *c;
466 os_cpu_mask_t cpu_mask;
470 memset(&cpu_mask, 0, sizeof(cpu_mask));
471 fio_cpu_set(&cpu_mask, t->cpu);
473 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
474 log_err("clock setaffinity failed\n");
478 pthread_mutex_lock(&t->lock);
479 pthread_mutex_unlock(&t->started);
483 for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
489 seq = atomic32_inc_return(t->seq);
492 tsc = get_cpu_clock();
493 } while (seq != *t->seq);
499 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
500 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
503 * The most common platform clock breakage is returning zero
504 * indefinitely. Check for that and return failure.
506 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
512 static int clock_cmp(const void *p1, const void *p2)
514 const struct clock_entry *c1 = p1;
515 const struct clock_entry *c2 = p2;
517 if (c1->seq == c2->seq)
518 log_err("cs: bug in atomic sequence!\n");
520 return c1->seq - c2->seq;
523 int fio_monotonic_clocktest(void)
525 struct clock_thread *cthreads;
526 unsigned int nr_cpus = cpus_online();
527 struct clock_entry *entries;
528 unsigned long tentries, failed = 0;
529 struct clock_entry *prev, *this;
533 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
536 fio_debug |= 1U << FD_TIME;
538 calibrate_cpu_clock();
540 fio_debug &= ~(1U << FD_TIME);
543 cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
544 tentries = CLOCK_ENTRIES * nr_cpus;
545 entries = malloc(tentries * sizeof(struct clock_entry));
547 log_info("cs: Testing %u CPUs\n", nr_cpus);
549 for (i = 0; i < nr_cpus; i++) {
550 struct clock_thread *t = &cthreads[i];
554 t->entries = &entries[i * CLOCK_ENTRIES];
555 pthread_mutex_init(&t->lock, NULL);
556 pthread_mutex_init(&t->started, NULL);
557 pthread_mutex_lock(&t->lock);
558 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
565 for (i = 0; i < nr_cpus; i++) {
566 struct clock_thread *t = &cthreads[i];
568 pthread_mutex_lock(&t->started);
571 for (i = 0; i < nr_cpus; i++) {
572 struct clock_thread *t = &cthreads[i];
574 pthread_mutex_unlock(&t->lock);
577 for (i = 0; i < nr_cpus; i++) {
578 struct clock_thread *t = &cthreads[i];
581 pthread_join(t->thread, &ret);
588 log_err("Clocksource test: %lu threads failed\n", failed);
592 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
594 for (failed = i = 0; i < tentries; i++) {
602 if (prev->tsc > this->tsc) {
603 uint64_t diff = prev->tsc - this->tsc;
605 log_info("cs: CPU clock mismatch (diff=%llu):\n",
606 (unsigned long long) diff);
607 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
608 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
616 log_info("cs: Failed: %lu\n", failed);
618 log_info("cs: Pass!\n");
625 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
627 int fio_monotonic_clocktest(void)
629 log_info("cs: current platform does not support CPU clocks\n");