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_gettime_offload(tp)))
216 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
217 static unsigned long get_cycles_per_usec(void)
221 enum fio_cs old_cs = fio_clock_source;
223 #ifdef CONFIG_CLOCK_GETTIME
224 fio_clock_source = CS_CGETTIME;
226 fio_clock_source = CS_GTOD;
230 c_s = get_cpu_clock();
236 elapsed = utime_since(&s, &e);
237 if (elapsed >= 1280) {
238 c_e = get_cpu_clock();
243 fio_clock_source = old_cs;
244 return (c_e - c_s + 127) >> 7;
247 #define NR_TIME_ITERS 50
249 static int calibrate_cpu_clock(void)
251 double delta, mean, S;
252 uint64_t avg, cycles[NR_TIME_ITERS];
255 cycles[0] = get_cycles_per_usec();
256 S = delta = mean = 0.0;
257 for (i = 0; i < NR_TIME_ITERS; i++) {
258 cycles[i] = get_cycles_per_usec();
259 delta = cycles[i] - mean;
261 mean += delta / (i + 1.0);
262 S += delta * (cycles[i] - mean);
267 * The most common platform clock breakage is returning zero
268 * indefinitely. Check for that and return failure.
270 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
273 S = sqrt(S / (NR_TIME_ITERS - 1.0));
276 for (i = 0; i < NR_TIME_ITERS; i++) {
277 double this = cycles[i];
279 if ((fmax(this, mean) - fmin(this, mean)) > S)
285 S /= (double) NR_TIME_ITERS;
288 for (i = 0; i < NR_TIME_ITERS; i++)
289 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
290 (unsigned long long) cycles[i] / 10);
293 avg = (avg + 5) / 10;
294 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
295 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
297 cycles_per_usec = avg;
298 inv_cycles_per_usec = 16777216UL / cycles_per_usec;
299 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
303 static int calibrate_cpu_clock(void)
305 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
311 #endif // ARCH_HAVE_CPU_CLOCK
313 #ifndef CONFIG_TLS_THREAD
314 void fio_local_clock_init(int is_thread)
318 t = calloc(1, sizeof(*t));
319 if (pthread_setspecific(tv_tls_key, t)) {
320 log_err("fio: can't set TLS key\n");
325 static void kill_tv_tls_key(void *data)
330 void fio_local_clock_init(int is_thread)
335 void fio_clock_init(void)
337 if (fio_clock_source == fio_clock_source_inited)
340 #ifndef CONFIG_TLS_THREAD
341 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
342 log_err("fio: can't create TLS key\n");
345 fio_clock_source_inited = fio_clock_source;
347 if (calibrate_cpu_clock())
351 * If the arch sets tsc_reliable != 0, then it must be good enough
352 * to use as THE clock source. For x86 CPUs, this means the TSC
353 * runs at a constant rate and is synced across CPU cores.
356 if (!fio_clock_source_set)
357 fio_clock_source = CS_CPUCLOCK;
358 } else if (fio_clock_source == CS_CPUCLOCK)
359 log_info("fio: clocksource=cpu may not be reliable\n");
362 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
367 sec = e->tv_sec - s->tv_sec;
368 usec = e->tv_usec - s->tv_usec;
369 if (sec > 0 && usec < 0) {
375 * time warp bug on some kernels?
377 if (sec < 0 || (sec == 0 && usec < 0))
380 ret = sec * 1000000ULL + usec;
385 uint64_t utime_since_now(const struct timeval *s)
389 fio_gettime(&t, NULL);
390 return utime_since(s, &t);
393 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
397 sec = e->tv_sec - s->tv_sec;
398 usec = e->tv_usec - s->tv_usec;
399 if (sec > 0 && usec < 0) {
404 if (sec < 0 || (sec == 0 && usec < 0))
414 uint64_t mtime_since_now(const struct timeval *s)
417 void *p = __builtin_return_address(0);
420 return mtime_since(s, &t);
423 uint64_t time_since_now(const struct timeval *s)
425 return mtime_since_now(s) / 1000;
428 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
431 #define CLOCK_ENTRIES 100000
439 struct clock_thread {
442 pthread_mutex_t lock;
443 pthread_mutex_t started;
445 struct clock_entry *entries;
448 static inline uint32_t atomic32_inc_return(uint32_t *seq)
450 return 1 + __sync_fetch_and_add(seq, 1);
453 static void *clock_thread_fn(void *data)
455 struct clock_thread *t = data;
456 struct clock_entry *c;
457 os_cpu_mask_t cpu_mask;
461 memset(&cpu_mask, 0, sizeof(cpu_mask));
462 fio_cpu_set(&cpu_mask, t->cpu);
464 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
465 log_err("clock setaffinity failed\n");
469 pthread_mutex_lock(&t->lock);
470 pthread_mutex_unlock(&t->started);
474 for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
480 seq = atomic32_inc_return(t->seq);
483 tsc = get_cpu_clock();
484 } while (seq != *t->seq);
490 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
491 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
494 * The most common platform clock breakage is returning zero
495 * indefinitely. Check for that and return failure.
497 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
503 static int clock_cmp(const void *p1, const void *p2)
505 const struct clock_entry *c1 = p1;
506 const struct clock_entry *c2 = p2;
508 if (c1->seq == c2->seq)
509 log_err("cs: bug in atomic sequence!\n");
511 return c1->seq - c2->seq;
514 int fio_monotonic_clocktest(void)
516 struct clock_thread *cthreads;
517 unsigned int nr_cpus = cpus_online();
518 struct clock_entry *entries;
519 unsigned long tentries, failed = 0;
520 struct clock_entry *prev, *this;
524 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
527 fio_debug |= 1U << FD_TIME;
529 calibrate_cpu_clock();
531 fio_debug &= ~(1U << FD_TIME);
534 cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
535 tentries = CLOCK_ENTRIES * nr_cpus;
536 entries = malloc(tentries * sizeof(struct clock_entry));
538 log_info("cs: Testing %u CPUs\n", nr_cpus);
540 for (i = 0; i < nr_cpus; i++) {
541 struct clock_thread *t = &cthreads[i];
545 t->entries = &entries[i * CLOCK_ENTRIES];
546 pthread_mutex_init(&t->lock, NULL);
547 pthread_mutex_init(&t->started, NULL);
548 pthread_mutex_lock(&t->lock);
549 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
556 for (i = 0; i < nr_cpus; i++) {
557 struct clock_thread *t = &cthreads[i];
559 pthread_mutex_lock(&t->started);
562 for (i = 0; i < nr_cpus; i++) {
563 struct clock_thread *t = &cthreads[i];
565 pthread_mutex_unlock(&t->lock);
568 for (i = 0; i < nr_cpus; i++) {
569 struct clock_thread *t = &cthreads[i];
572 pthread_join(t->thread, &ret);
579 log_err("Clocksource test: %lu threads failed\n", failed);
583 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
585 for (failed = i = 0; i < tentries; i++) {
593 if (prev->tsc > this->tsc) {
594 uint64_t diff = prev->tsc - this->tsc;
596 log_info("cs: CPU clock mismatch (diff=%llu):\n",
597 (unsigned long long) diff);
598 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
599 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
607 log_info("cs: Failed: %lu\n", failed);
609 log_info("cs: Pass!\n");
616 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
618 int fio_monotonic_clocktest(void)
620 log_info("cs: current platform does not support CPU clocks\n");