nanosecond: initial commit changing timeval to timespec
[fio.git] / gettime.c
... / ...
CommitLineData
1/*
2 * Clock functions
3 */
4
5#include <unistd.h>
6#include <math.h>
7#include <sys/time.h>
8#include <time.h>
9
10#include "fio.h"
11#include "smalloc.h"
12
13#include "hash.h"
14#include "os/os.h"
15
16#if defined(ARCH_HAVE_CPU_CLOCK)
17#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
18static unsigned long cycles_per_usec;
19static unsigned long inv_cycles_per_nsec;
20static uint64_t max_cycles_for_mult;
21#define NSEC_INV_FACTOR 4096
22#endif
23#ifdef ARCH_CPU_CLOCK_WRAPS
24static unsigned long long cycles_start, cycles_wrap;
25#endif
26#endif
27int tsc_reliable = 0;
28
29struct tv_valid {
30 uint64_t last_cycles;
31 int last_tv_valid;
32 int warned;
33};
34#ifdef ARCH_HAVE_CPU_CLOCK
35#ifdef CONFIG_TLS_THREAD
36static __thread struct tv_valid static_tv_valid;
37#else
38static pthread_key_t tv_tls_key;
39#endif
40#endif
41
42enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
43int fio_clock_source_set = 0;
44static enum fio_cs fio_clock_source_inited = CS_INVAL;
45
46#ifdef FIO_DEBUG_TIME
47
48#define HASH_BITS 8
49#define HASH_SIZE (1 << HASH_BITS)
50
51static struct flist_head hash[HASH_SIZE];
52static int gtod_inited;
53
54struct gtod_log {
55 struct flist_head list;
56 void *caller;
57 unsigned long calls;
58};
59
60static struct gtod_log *find_hash(void *caller)
61{
62 unsigned long h = hash_ptr(caller, HASH_BITS);
63 struct flist_head *entry;
64
65 flist_for_each(entry, &hash[h]) {
66 struct gtod_log *log = flist_entry(entry, struct gtod_log,
67 list);
68
69 if (log->caller == caller)
70 return log;
71 }
72
73 return NULL;
74}
75
76static void inc_caller(void *caller)
77{
78 struct gtod_log *log = find_hash(caller);
79
80 if (!log) {
81 unsigned long h;
82
83 log = malloc(sizeof(*log));
84 INIT_FLIST_HEAD(&log->list);
85 log->caller = caller;
86 log->calls = 0;
87
88 h = hash_ptr(caller, HASH_BITS);
89 flist_add_tail(&log->list, &hash[h]);
90 }
91
92 log->calls++;
93}
94
95static void gtod_log_caller(void *caller)
96{
97 if (gtod_inited)
98 inc_caller(caller);
99}
100
101static void fio_exit fio_dump_gtod(void)
102{
103 unsigned long total_calls = 0;
104 int i;
105
106 for (i = 0; i < HASH_SIZE; i++) {
107 struct flist_head *entry;
108 struct gtod_log *log;
109
110 flist_for_each(entry, &hash[i]) {
111 log = flist_entry(entry, struct gtod_log, list);
112
113 printf("function %p, calls %lu\n", log->caller,
114 log->calls);
115 total_calls += log->calls;
116 }
117 }
118
119 printf("Total %lu gettimeofday\n", total_calls);
120}
121
122static void fio_init gtod_init(void)
123{
124 int i;
125
126 for (i = 0; i < HASH_SIZE; i++)
127 INIT_FLIST_HEAD(&hash[i]);
128
129 gtod_inited = 1;
130}
131
132#endif /* FIO_DEBUG_TIME */
133
134#ifdef CONFIG_CLOCK_GETTIME
135static int fill_clock_gettime(struct timespec *ts)
136{
137#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
138 return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
139#elif defined(CONFIG_CLOCK_MONOTONIC)
140 return clock_gettime(CLOCK_MONOTONIC, ts);
141#else
142 return clock_gettime(CLOCK_REALTIME, ts);
143#endif
144}
145#endif
146
147static void __fio_gettime(struct timespec *tp)
148{
149 switch (fio_clock_source) {
150#ifdef CONFIG_GETTIMEOFDAY
151 case CS_GTOD: {
152 struct timeval tv;
153 gettimeofday(&tv, NULL);
154
155 tp->tv_sec = tv.tv_sec;
156 tp->tv_nsec = tv.tv_usec * 1000;
157 break;
158 }
159#endif
160#ifdef CONFIG_CLOCK_GETTIME
161 case CS_CGETTIME: {
162 if (fill_clock_gettime(tp) < 0) {
163 log_err("fio: clock_gettime fails\n");
164 assert(0);
165 }
166 break;
167 }
168#endif
169#ifdef ARCH_HAVE_CPU_CLOCK
170 case CS_CPUCLOCK: {
171 uint64_t nsecs, t;
172 struct tv_valid *tv;
173
174#ifdef CONFIG_TLS_THREAD
175 tv = &static_tv_valid;
176#else
177 tv = pthread_getspecific(tv_tls_key);
178#endif
179
180 t = get_cpu_clock();
181#ifdef ARCH_CPU_CLOCK_WRAPS
182 if (t < cycles_start && !cycles_wrap)
183 cycles_wrap = 1;
184 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
185 log_err("fio: double CPU clock wrap\n");
186 tv->warned = 1;
187 }
188
189 t -= cycles_start;
190#endif
191 tv->last_cycles = t;
192 tv->last_tv_valid = 1;
193#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
194 nsecs = t * 1000 / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
195#else
196 if (t < max_cycles_for_mult)
197 nsecs = (t * inv_cycles_per_nsec) / NSEC_INV_FACTOR;
198 else
199 nsecs = (t / NSEC_INV_FACTOR) * inv_cycles_per_nsec;
200#endif
201 tp->tv_sec = nsecs / 1000000000ULL;
202 tp->tv_nsec = nsecs % 1000000000ULL;
203 break;
204 }
205#endif
206 default:
207 log_err("fio: invalid clock source %d\n", fio_clock_source);
208 break;
209 }
210}
211
212#ifdef FIO_DEBUG_TIME
213void fio_gettime(struct timespec *tp, void *caller)
214#else
215void fio_gettime(struct timespec *tp, void fio_unused *caller)
216#endif
217{
218#ifdef FIO_DEBUG_TIME
219 if (!caller)
220 caller = __builtin_return_address(0);
221
222 gtod_log_caller(caller);
223#endif
224 if (fio_unlikely(fio_gettime_offload(tp)))
225 return;
226
227 __fio_gettime(tp);
228}
229
230#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
231static unsigned long get_cycles_per_usec(void)
232{
233 struct timespec s, e;
234 uint64_t c_s, c_e;
235 enum fio_cs old_cs = fio_clock_source;
236 uint64_t elapsed;
237
238#ifdef CONFIG_CLOCK_GETTIME
239 fio_clock_source = CS_CGETTIME;
240#else
241 fio_clock_source = CS_GTOD;
242#endif
243 __fio_gettime(&s);
244
245 c_s = get_cpu_clock();
246 do {
247 __fio_gettime(&e);
248
249 elapsed = utime_since(&s, &e);
250 if (elapsed >= 1280) {
251 c_e = get_cpu_clock();
252 break;
253 }
254 } while (1);
255
256 fio_clock_source = old_cs;
257 return (c_e - c_s) / elapsed;
258}
259
260#define NR_TIME_ITERS 50
261
262static int calibrate_cpu_clock(void)
263{
264 double delta, mean, S;
265 uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
266 int i, samples;
267
268 cycles[0] = get_cycles_per_usec();
269 S = delta = mean = 0.0;
270 for (i = 0; i < NR_TIME_ITERS; i++) {
271 cycles[i] = get_cycles_per_usec();
272 delta = cycles[i] - mean;
273 if (delta) {
274 mean += delta / (i + 1.0);
275 S += delta * (cycles[i] - mean);
276 }
277 }
278
279 /*
280 * The most common platform clock breakage is returning zero
281 * indefinitely. Check for that and return failure.
282 */
283 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
284 return 1;
285
286 S = sqrt(S / (NR_TIME_ITERS - 1.0));
287
288 minc = -1ULL;
289 maxc = samples = avg = 0;
290 for (i = 0; i < NR_TIME_ITERS; i++) {
291 double this = cycles[i];
292
293 minc = min(cycles[i], minc);
294 maxc = max(cycles[i], maxc);
295
296 if ((fmax(this, mean) - fmin(this, mean)) > S)
297 continue;
298 samples++;
299 avg += this;
300 }
301
302 S /= (double) NR_TIME_ITERS;
303
304 for (i = 0; i < NR_TIME_ITERS; i++)
305 dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
306
307 avg /= samples;
308 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
309 dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
310 (unsigned long long) minc,
311 (unsigned long long) maxc, mean, S);
312
313 cycles_per_usec = avg;
314 inv_cycles_per_nsec = NSEC_INV_FACTOR * 1000 / cycles_per_usec;
315 max_cycles_for_mult = ~0ULL / inv_cycles_per_nsec;
316 dprint(FD_TIME, "inv_cycles_per_nsec=%lu\n", inv_cycles_per_nsec);
317#ifdef ARCH_CPU_CLOCK_WRAPS
318 cycles_start = get_cpu_clock();
319 dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
320#endif
321 return 0;
322}
323#else
324static int calibrate_cpu_clock(void)
325{
326#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
327 return 0;
328#else
329 return 1;
330#endif
331}
332#endif // ARCH_HAVE_CPU_CLOCK
333
334#ifndef CONFIG_TLS_THREAD
335void fio_local_clock_init(int is_thread)
336{
337 struct tv_valid *t;
338
339 t = calloc(1, sizeof(*t));
340 if (pthread_setspecific(tv_tls_key, t)) {
341 log_err("fio: can't set TLS key\n");
342 assert(0);
343 }
344}
345
346static void kill_tv_tls_key(void *data)
347{
348 free(data);
349}
350#else
351void fio_local_clock_init(int is_thread)
352{
353}
354#endif
355
356void fio_clock_init(void)
357{
358 if (fio_clock_source == fio_clock_source_inited)
359 return;
360
361#ifndef CONFIG_TLS_THREAD
362 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
363 log_err("fio: can't create TLS key\n");
364#endif
365
366 fio_clock_source_inited = fio_clock_source;
367
368 if (calibrate_cpu_clock())
369 tsc_reliable = 0;
370
371 /*
372 * If the arch sets tsc_reliable != 0, then it must be good enough
373 * to use as THE clock source. For x86 CPUs, this means the TSC
374 * runs at a constant rate and is synced across CPU cores.
375 */
376 if (tsc_reliable) {
377 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
378 fio_clock_source = CS_CPUCLOCK;
379 } else if (fio_clock_source == CS_CPUCLOCK)
380 log_info("fio: clocksource=cpu may not be reliable\n");
381}
382
383uint64_t utime_since(const struct timespec *s, const struct timespec *e)
384{
385 int64_t sec, usec;
386
387 sec = e->tv_sec - s->tv_sec;
388 usec = (e->tv_nsec - s->tv_nsec) / 1000;
389 if (sec > 0 && usec < 0) {
390 sec--;
391 usec += 1000000;
392 }
393
394 /*
395 * time warp bug on some kernels?
396 */
397 if (sec < 0 || (sec == 0 && usec < 0))
398 return 0;
399
400 return usec + (sec * 1000000);
401}
402
403uint64_t utime_since_now(const struct timespec *s)
404{
405 struct timespec t;
406#ifdef FIO_DEBUG_TIME
407 void *p = __builtin_return_address(0);
408
409 fio_gettime(&t, p);
410#else
411 fio_gettime(&t, NULL);
412#endif
413
414 return utime_since(s, &t);
415}
416
417uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
418{
419 int64_t sec, usec;
420
421 sec = e->tv_sec - s->tv_sec;
422 usec = (e->tv_usec - s->tv_usec);
423 if (sec > 0 && usec < 0) {
424 sec--;
425 usec += 1000000;
426 }
427
428 if (sec < 0 || (sec == 0 && usec < 0))
429 return 0;
430
431 sec *= 1000;
432 usec /= 1000;
433 return sec + usec;
434}
435
436uint64_t mtime_since_now(const struct timespec *s)
437{
438 struct timespec t;
439#ifdef FIO_DEBUG_TIME
440 void *p = __builtin_return_address(0);
441
442 fio_gettime(&t, p);
443#else
444 fio_gettime(&t, NULL);
445#endif
446
447 return mtime_since(s, &t);
448}
449
450uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
451{
452 int64_t sec, usec;
453
454 sec = e->tv_sec - s->tv_sec;
455 usec = (e->tv_nsec - s->tv_nsec) / 1000;
456 if (sec > 0 && usec < 0) {
457 sec--;
458 usec += 1000000;
459 }
460
461 if (sec < 0 || (sec == 0 && usec < 0))
462 return 0;
463
464 sec *= 1000;
465 usec /= 1000;
466 return sec + usec;
467}
468
469uint64_t time_since_now(const struct timespec *s)
470{
471 return mtime_since_now(s) / 1000;
472}
473
474#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
475 defined(CONFIG_SFAA)
476
477#define CLOCK_ENTRIES_DEBUG 100000
478#define CLOCK_ENTRIES_TEST 10000
479
480struct clock_entry {
481 uint32_t seq;
482 uint32_t cpu;
483 uint64_t tsc;
484};
485
486struct clock_thread {
487 pthread_t thread;
488 int cpu;
489 int debug;
490 pthread_mutex_t lock;
491 pthread_mutex_t started;
492 unsigned long nr_entries;
493 uint32_t *seq;
494 struct clock_entry *entries;
495};
496
497static inline uint32_t atomic32_inc_return(uint32_t *seq)
498{
499 return 1 + __sync_fetch_and_add(seq, 1);
500}
501
502static void *clock_thread_fn(void *data)
503{
504 struct clock_thread *t = data;
505 struct clock_entry *c;
506 os_cpu_mask_t cpu_mask;
507 uint32_t last_seq;
508 unsigned long long first;
509 int i;
510
511 if (fio_cpuset_init(&cpu_mask)) {
512 int __err = errno;
513
514 log_err("clock cpuset init failed: %s\n", strerror(__err));
515 goto err_out;
516 }
517
518 fio_cpu_set(&cpu_mask, t->cpu);
519
520 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
521 int __err = errno;
522
523 log_err("clock setaffinity failed: %s\n", strerror(__err));
524 goto err;
525 }
526
527 pthread_mutex_lock(&t->lock);
528 pthread_mutex_unlock(&t->started);
529
530 first = get_cpu_clock();
531 last_seq = 0;
532 c = &t->entries[0];
533 for (i = 0; i < t->nr_entries; i++, c++) {
534 uint32_t seq;
535 uint64_t tsc;
536
537 c->cpu = t->cpu;
538 do {
539 seq = atomic32_inc_return(t->seq);
540 if (seq < last_seq)
541 break;
542 tsc = get_cpu_clock();
543 } while (seq != *t->seq);
544
545 c->seq = seq;
546 c->tsc = tsc;
547 }
548
549 if (t->debug) {
550 unsigned long long clocks;
551
552 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
553 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
554 clocks, first);
555 }
556
557 /*
558 * The most common platform clock breakage is returning zero
559 * indefinitely. Check for that and return failure.
560 */
561 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
562 goto err;
563
564 fio_cpuset_exit(&cpu_mask);
565 return NULL;
566err:
567 fio_cpuset_exit(&cpu_mask);
568err_out:
569 return (void *) 1;
570}
571
572static int clock_cmp(const void *p1, const void *p2)
573{
574 const struct clock_entry *c1 = p1;
575 const struct clock_entry *c2 = p2;
576
577 if (c1->seq == c2->seq)
578 log_err("cs: bug in atomic sequence!\n");
579
580 return c1->seq - c2->seq;
581}
582
583int fio_monotonic_clocktest(int debug)
584{
585 struct clock_thread *cthreads;
586 unsigned int nr_cpus = cpus_online();
587 struct clock_entry *entries;
588 unsigned long nr_entries, tentries, failed = 0;
589 struct clock_entry *prev, *this;
590 uint32_t seq = 0;
591 unsigned int i;
592
593 if (debug) {
594 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
595
596#ifdef FIO_INC_DEBUG
597 fio_debug |= 1U << FD_TIME;
598#endif
599 nr_entries = CLOCK_ENTRIES_DEBUG;
600 } else
601 nr_entries = CLOCK_ENTRIES_TEST;
602
603 calibrate_cpu_clock();
604
605 if (debug) {
606#ifdef FIO_INC_DEBUG
607 fio_debug &= ~(1U << FD_TIME);
608#endif
609 }
610
611 cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
612 tentries = nr_entries * nr_cpus;
613 entries = malloc(tentries * sizeof(struct clock_entry));
614
615 if (debug)
616 log_info("cs: Testing %u CPUs\n", nr_cpus);
617
618 for (i = 0; i < nr_cpus; i++) {
619 struct clock_thread *t = &cthreads[i];
620
621 t->cpu = i;
622 t->debug = debug;
623 t->seq = &seq;
624 t->nr_entries = nr_entries;
625 t->entries = &entries[i * nr_entries];
626 pthread_mutex_init(&t->lock, NULL);
627 pthread_mutex_init(&t->started, NULL);
628 pthread_mutex_lock(&t->lock);
629 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
630 failed++;
631 nr_cpus = i;
632 break;
633 }
634 }
635
636 for (i = 0; i < nr_cpus; i++) {
637 struct clock_thread *t = &cthreads[i];
638
639 pthread_mutex_lock(&t->started);
640 }
641
642 for (i = 0; i < nr_cpus; i++) {
643 struct clock_thread *t = &cthreads[i];
644
645 pthread_mutex_unlock(&t->lock);
646 }
647
648 for (i = 0; i < nr_cpus; i++) {
649 struct clock_thread *t = &cthreads[i];
650 void *ret;
651
652 pthread_join(t->thread, &ret);
653 if (ret)
654 failed++;
655 }
656 free(cthreads);
657
658 if (failed) {
659 if (debug)
660 log_err("Clocksource test: %lu threads failed\n", failed);
661 goto err;
662 }
663
664 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
665
666 /* silence silly gcc */
667 prev = NULL;
668 for (failed = i = 0; i < tentries; i++) {
669 this = &entries[i];
670
671 if (!i) {
672 prev = this;
673 continue;
674 }
675
676 if (prev->tsc > this->tsc) {
677 uint64_t diff = prev->tsc - this->tsc;
678
679 if (!debug) {
680 failed++;
681 break;
682 }
683
684 log_info("cs: CPU clock mismatch (diff=%llu):\n",
685 (unsigned long long) diff);
686 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
687 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
688 failed++;
689 }
690
691 prev = this;
692 }
693
694 if (debug) {
695 if (failed)
696 log_info("cs: Failed: %lu\n", failed);
697 else
698 log_info("cs: Pass!\n");
699 }
700err:
701 free(entries);
702 return !!failed;
703}
704
705#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
706
707int fio_monotonic_clocktest(int debug)
708{
709 if (debug)
710 log_info("cs: current platform does not support CPU clocks\n");
711 return 1;
712}
713
714#endif