engines/libaio: fix issue with EAGAIN
[fio.git] / gettime.c
CommitLineData
02bcaa8c 1/*
f5cc024a 2 * Clock functions
02bcaa8c 3 */
f5cc024a 4
02bcaa8c 5#include <unistd.h>
c223da83 6#include <math.h>
02bcaa8c 7#include <sys/time.h>
03e20d68 8#include <time.h>
02bcaa8c
JA
9
10#include "fio.h"
be4ecfdf 11#include "smalloc.h"
02bcaa8c
JA
12
13#include "hash.h"
7d11f871 14#include "os/os.h"
02bcaa8c 15
919e789d 16#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
c223da83 17static unsigned long cycles_per_usec;
71339117 18static unsigned long inv_cycles_per_usec;
09a32402 19#endif
4de98eb0 20int tsc_reliable = 0;
5d879392
JA
21
22struct tv_valid {
23 struct timeval last_tv;
ba458c2f 24 uint64_t last_cycles;
5d879392
JA
25 int last_tv_valid;
26};
67bf9823 27#ifdef CONFIG_TLS_THREAD
b4ea84da 28static __thread struct tv_valid static_tv_valid;
67bf9823 29#else
5d879392 30static pthread_key_t tv_tls_key;
67bf9823 31#endif
02bcaa8c 32
16de1bf9 33enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
fa80feae 34int fio_clock_source_set = 0;
10aa136b 35static enum fio_cs fio_clock_source_inited = CS_INVAL;
c223da83 36
02bcaa8c
JA
37#ifdef FIO_DEBUG_TIME
38
39#define HASH_BITS 8
40#define HASH_SIZE (1 << HASH_BITS)
41
01743ee1 42static struct flist_head hash[HASH_SIZE];
02bcaa8c
JA
43static int gtod_inited;
44
45struct gtod_log {
01743ee1 46 struct flist_head list;
02bcaa8c
JA
47 void *caller;
48 unsigned long calls;
49};
50
51static struct gtod_log *find_hash(void *caller)
52{
53 unsigned long h = hash_ptr(caller, HASH_BITS);
01743ee1 54 struct flist_head *entry;
02bcaa8c 55
01743ee1
JA
56 flist_for_each(entry, &hash[h]) {
57 struct gtod_log *log = flist_entry(entry, struct gtod_log,
58 list);
02bcaa8c
JA
59
60 if (log->caller == caller)
61 return log;
62 }
63
64 return NULL;
65}
66
67static struct gtod_log *find_log(void *caller)
68{
69 struct gtod_log *log = find_hash(caller);
70
71 if (!log) {
72 unsigned long h;
73
74 log = malloc(sizeof(*log));
01743ee1 75 INIT_FLIST_HEAD(&log->list);
02bcaa8c
JA
76 log->caller = caller;
77 log->calls = 0;
78
79 h = hash_ptr(caller, HASH_BITS);
01743ee1 80 flist_add_tail(&log->list, &hash[h]);
02bcaa8c
JA
81 }
82
83 return log;
84}
85
86static void gtod_log_caller(void *caller)
87{
88 if (gtod_inited) {
89 struct gtod_log *log = find_log(caller);
90
91 log->calls++;
92 }
93}
94
95static void fio_exit fio_dump_gtod(void)
96{
97 unsigned long total_calls = 0;
98 int i;
99
100 for (i = 0; i < HASH_SIZE; i++) {
01743ee1 101 struct flist_head *entry;
02bcaa8c
JA
102 struct gtod_log *log;
103
01743ee1
JA
104 flist_for_each(entry, &hash[i]) {
105 log = flist_entry(entry, struct gtod_log, list);
02bcaa8c 106
5ec10eaa
JA
107 printf("function %p, calls %lu\n", log->caller,
108 log->calls);
02bcaa8c
JA
109 total_calls += log->calls;
110 }
111 }
112
113 printf("Total %lu gettimeofday\n", total_calls);
114}
115
116static void fio_init gtod_init(void)
117{
118 int i;
119
120 for (i = 0; i < HASH_SIZE; i++)
01743ee1 121 INIT_FLIST_HEAD(&hash[i]);
02bcaa8c
JA
122
123 gtod_inited = 1;
124}
125
126#endif /* FIO_DEBUG_TIME */
127
67bf9823 128#ifdef CONFIG_CLOCK_GETTIME
9ff1c070
JA
129static int fill_clock_gettime(struct timespec *ts)
130{
67bf9823 131#ifdef CONFIG_CLOCK_MONOTONIC
9ff1c070
JA
132 return clock_gettime(CLOCK_MONOTONIC, ts);
133#else
134 return clock_gettime(CLOCK_REALTIME, ts);
135#endif
136}
1e97cce9 137#endif
67bf9823
JA
138
139static void *__fio_gettime(struct timeval *tp)
02bcaa8c 140{
93f0b09a 141 struct tv_valid *tv;
5d879392 142
67bf9823
JA
143#ifdef CONFIG_TLS_THREAD
144 tv = &static_tv_valid;
145#else
93f0b09a 146 tv = pthread_getspecific(tv_tls_key);
67bf9823 147#endif
93f0b09a 148
c223da83 149 switch (fio_clock_source) {
67bf9823 150#ifdef CONFIG_GETTIMEOFDAY
c223da83 151 case CS_GTOD:
02bcaa8c 152 gettimeofday(tp, NULL);
c223da83 153 break;
67bf9823
JA
154#endif
155#ifdef CONFIG_CLOCK_GETTIME
c223da83 156 case CS_CGETTIME: {
02bcaa8c
JA
157 struct timespec ts;
158
9ff1c070 159 if (fill_clock_gettime(&ts) < 0) {
c223da83
JA
160 log_err("fio: clock_gettime fails\n");
161 assert(0);
02bcaa8c
JA
162 }
163
164 tp->tv_sec = ts.tv_sec;
165 tp->tv_usec = ts.tv_nsec / 1000;
c223da83
JA
166 break;
167 }
67bf9823 168#endif
c223da83
JA
169#ifdef ARCH_HAVE_CPU_CLOCK
170 case CS_CPUCLOCK: {
ba458c2f 171 uint64_t usecs, t;
c223da83
JA
172
173 t = get_cpu_clock();
93f0b09a 174 if (tv && t < tv->last_cycles) {
c223da83 175 dprint(FD_TIME, "CPU clock going back in time\n");
93f0b09a
JA
176 t = tv->last_cycles;
177 } else if (tv)
178 tv->last_cycles = t;
c223da83 179
919e789d
CE
180#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
181 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
182#else
71339117 183 usecs = (t * inv_cycles_per_usec) / 16777216UL;
919e789d 184#endif
c223da83
JA
185 tp->tv_sec = usecs / 1000000;
186 tp->tv_usec = usecs % 1000000;
c223da83
JA
187 break;
188 }
189#endif
190 default:
191 log_err("fio: invalid clock source %d\n", fio_clock_source);
192 break;
02bcaa8c 193 }
3e488920 194
67bf9823
JA
195 return tv;
196}
197
198#ifdef FIO_DEBUG_TIME
199void fio_gettime(struct timeval *tp, void *caller)
200#else
201void fio_gettime(struct timeval *tp, void fio_unused *caller)
202#endif
203{
204 struct tv_valid *tv;
205
206#ifdef FIO_DEBUG_TIME
207 if (!caller)
208 caller = __builtin_return_address(0);
209
210 gtod_log_caller(caller);
211#endif
225ba9e3 212 if (fio_unlikely(fio_tv)) {
67bf9823
JA
213 memcpy(tp, fio_tv, sizeof(*tp));
214 return;
215 }
216
217 tv = __fio_gettime(tp);
218
3e488920
JA
219 /*
220 * If Linux is using the tsc clock on non-synced processors,
221 * sometimes time can appear to drift backwards. Fix that up.
222 */
93f0b09a
JA
223 if (tv) {
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;
5d879392 230 }
93f0b09a
JA
231 tv->last_tv_valid = 1;
232 memcpy(&tv->last_tv, tp, sizeof(*tp));
3e488920 233 }
02bcaa8c 234}
be4ecfdf 235
919e789d 236#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
c223da83
JA
237static unsigned long get_cycles_per_usec(void)
238{
239 struct timeval s, e;
ba458c2f 240 uint64_t c_s, c_e;
67bf9823 241 enum fio_cs old_cs = fio_clock_source;
c223da83 242
67bf9823
JA
243#ifdef CONFIG_CLOCK_GETTIME
244 fio_clock_source = CS_CGETTIME;
245#else
246 fio_clock_source = CS_GTOD;
247#endif
248 __fio_gettime(&s);
9ff1c070 249
c223da83
JA
250 c_s = get_cpu_clock();
251 do {
ba458c2f 252 uint64_t elapsed;
c223da83 253
67bf9823 254 __fio_gettime(&e);
9ff1c070 255
c223da83 256 elapsed = utime_since(&s, &e);
486332e5 257 if (elapsed >= 1280) {
c223da83
JA
258 c_e = get_cpu_clock();
259 break;
260 }
261 } while (1);
262
67bf9823 263 fio_clock_source = old_cs;
c011000d 264 return (c_e - c_s + 127) >> 7;
c223da83
JA
265}
266
fa80feae
JA
267#define NR_TIME_ITERS 50
268
e259879e 269static int calibrate_cpu_clock(void)
c223da83
JA
270{
271 double delta, mean, S;
ba458c2f 272 uint64_t avg, cycles[NR_TIME_ITERS];
c223da83
JA
273 int i, samples;
274
c223da83
JA
275 cycles[0] = get_cycles_per_usec();
276 S = delta = mean = 0.0;
fa80feae 277 for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83
JA
278 cycles[i] = get_cycles_per_usec();
279 delta = cycles[i] - mean;
280 if (delta) {
281 mean += delta / (i + 1.0);
282 S += delta * (cycles[i] - mean);
283 }
284 }
285
e259879e
JA
286 /*
287 * The most common platform clock breakage is returning zero
288 * indefinitely. Check for that and return failure.
289 */
290 if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
291 return 1;
292
fa80feae 293 S = sqrt(S / (NR_TIME_ITERS - 1.0));
c223da83
JA
294
295 samples = avg = 0;
fa80feae 296 for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83
JA
297 double this = cycles[i];
298
03e20d68 299 if ((fmax(this, mean) - fmin(this, mean)) > S)
c223da83
JA
300 continue;
301 samples++;
302 avg += this;
303 }
304
fa80feae 305 S /= (double) NR_TIME_ITERS;
89db727d 306 mean /= 10.0;
c223da83 307
fa80feae 308 for (i = 0; i < NR_TIME_ITERS; i++)
4b91ee8f
JA
309 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
310 (unsigned long long) cycles[i] / 10);
c223da83 311
d7abad3d 312 avg /= samples;
b0ff22d7 313 avg = (avg + 5) / 10;
4b91ee8f 314 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
c223da83
JA
315 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
316
317 cycles_per_usec = avg;
71339117
JA
318 inv_cycles_per_usec = 16777216UL / cycles_per_usec;
319 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
e259879e 320 return 0;
09a32402
JA
321}
322#else
e259879e 323static int calibrate_cpu_clock(void)
09a32402 324{
919e789d
CE
325#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
326 return 0;
327#else
e259879e 328 return 1;
09a32402 329#endif
919e789d
CE
330}
331#endif // ARCH_HAVE_CPU_CLOCK
09a32402 332
67bf9823 333#ifndef CONFIG_TLS_THREAD
5d879392
JA
334void fio_local_clock_init(int is_thread)
335{
336 struct tv_valid *t;
337
572cfb3f 338 t = calloc(1, sizeof(*t));
5d879392
JA
339 if (pthread_setspecific(tv_tls_key, t))
340 log_err("fio: can't set TLS key\n");
341}
342
343static void kill_tv_tls_key(void *data)
344{
345 free(data);
346}
67bf9823
JA
347#else
348void fio_local_clock_init(int is_thread)
349{
350}
351#endif
5d879392 352
09a32402
JA
353void fio_clock_init(void)
354{
01423eae
JA
355 if (fio_clock_source == fio_clock_source_inited)
356 return;
357
67bf9823 358#ifndef CONFIG_TLS_THREAD
5d879392
JA
359 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
360 log_err("fio: can't create TLS key\n");
67bf9823 361#endif
5d879392 362
01423eae 363 fio_clock_source_inited = fio_clock_source;
e259879e
JA
364
365 if (calibrate_cpu_clock())
366 tsc_reliable = 0;
fa80feae
JA
367
368 /*
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.
372 */
373 if (tsc_reliable) {
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");
c223da83
JA
378}
379
aa60bc58 380uint64_t utime_since(struct timeval *s, struct timeval *e)
be4ecfdf 381{
39ab7da2 382 long sec, usec;
aa60bc58 383 uint64_t ret;
39ab7da2
JA
384
385 sec = e->tv_sec - s->tv_sec;
386 usec = e->tv_usec - s->tv_usec;
387 if (sec > 0 && usec < 0) {
388 sec--;
389 usec += 1000000;
390 }
391
392 /*
393 * time warp bug on some kernels?
394 */
395 if (sec < 0 || (sec == 0 && usec < 0))
396 return 0;
397
398 ret = sec * 1000000ULL + usec;
399
400 return ret;
be4ecfdf
JA
401}
402
aa60bc58 403uint64_t utime_since_now(struct timeval *s)
be4ecfdf 404{
39ab7da2
JA
405 struct timeval t;
406
407 fio_gettime(&t, NULL);
408 return utime_since(s, &t);
be4ecfdf 409}
783a3eb1 410
aa60bc58 411uint64_t mtime_since(struct timeval *s, struct timeval *e)
783a3eb1 412{
39ab7da2 413 long sec, usec, ret;
783a3eb1 414
39ab7da2
JA
415 sec = e->tv_sec - s->tv_sec;
416 usec = e->tv_usec - s->tv_usec;
417 if (sec > 0 && usec < 0) {
418 sec--;
419 usec += 1000000;
783a3eb1
JA
420 }
421
39ab7da2
JA
422 if (sec < 0 || (sec == 0 && usec < 0))
423 return 0;
424
425 sec *= 1000UL;
426 usec /= 1000UL;
427 ret = sec + usec;
428
429 return ret;
783a3eb1
JA
430}
431
aa60bc58 432uint64_t mtime_since_now(struct timeval *s)
783a3eb1 433{
39ab7da2
JA
434 struct timeval t;
435 void *p = __builtin_return_address(0);
783a3eb1 436
39ab7da2
JA
437 fio_gettime(&t, p);
438 return mtime_since(s, &t);
439}
783a3eb1 440
aa60bc58 441uint64_t time_since_now(struct timeval *s)
39ab7da2
JA
442{
443 return mtime_since_now(s) / 1000;
783a3eb1 444}
7d11f871 445
67bf9823
JA
446#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
447 defined(CONFIG_SFAA)
7d11f871
JA
448
449#define CLOCK_ENTRIES 100000
450
451struct clock_entry {
58002f9a
JA
452 uint32_t seq;
453 uint32_t cpu;
ba458c2f 454 uint64_t tsc;
7d11f871
JA
455};
456
457struct clock_thread {
458 pthread_t thread;
459 int cpu;
460 pthread_mutex_t lock;
461 pthread_mutex_t started;
58002f9a 462 uint32_t *seq;
7d11f871
JA
463 struct clock_entry *entries;
464};
465
58002f9a 466static inline uint32_t atomic32_inc_return(uint32_t *seq)
7d11f871
JA
467{
468 return 1 + __sync_fetch_and_add(seq, 1);
469}
470
471static void *clock_thread_fn(void *data)
472{
473 struct clock_thread *t = data;
474 struct clock_entry *c;
475 os_cpu_mask_t cpu_mask;
58002f9a 476 uint32_t last_seq;
7d11f871
JA
477 int i;
478
479 memset(&cpu_mask, 0, sizeof(cpu_mask));
480 fio_cpu_set(&cpu_mask, t->cpu);
481
482 if (fio_setaffinity(gettid(), cpu_mask) == -1) {
483 log_err("clock setaffinity failed\n");
484 return (void *) 1;
485 }
486
7d11f871 487 pthread_mutex_lock(&t->lock);
b9b3498e 488 pthread_mutex_unlock(&t->started);
7d11f871 489
58002f9a 490 last_seq = 0;
7d11f871
JA
491 c = &t->entries[0];
492 for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
58002f9a
JA
493 uint32_t seq;
494 uint64_t tsc;
7d11f871
JA
495
496 c->cpu = t->cpu;
497 do {
58002f9a
JA
498 seq = atomic32_inc_return(t->seq);
499 if (seq < last_seq)
500 break;
7d11f871
JA
501 tsc = get_cpu_clock();
502 } while (seq != *t->seq);
503
504 c->seq = seq;
505 c->tsc = tsc;
506 }
507
0f78b22c
JA
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);
58002f9a 510
e259879e
JA
511 /*
512 * The most common platform clock breakage is returning zero
513 * indefinitely. Check for that and return failure.
514 */
58002f9a 515 if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
e259879e
JA
516 return (void *) 1;
517
7d11f871
JA
518 return NULL;
519}
520
521static int clock_cmp(const void *p1, const void *p2)
522{
523 const struct clock_entry *c1 = p1;
524 const struct clock_entry *c2 = p2;
525
b9b3498e
JA
526 if (c1->seq == c2->seq)
527 log_err("cs: bug in atomic sequence!\n");
528
7d11f871
JA
529 return c1->seq - c2->seq;
530}
531
532int fio_monotonic_clocktest(void)
533{
534 struct clock_thread *threads;
535 unsigned int nr_cpus = cpus_online();
536 struct clock_entry *entries;
814917be 537 unsigned long tentries, failed = 0;
80da8a8f 538 struct clock_entry *prev, *this;
58002f9a 539 uint32_t seq = 0;
caa3eb1c 540 unsigned int i;
7d11f871 541
d5e3f5d8
JA
542 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
543
b5b571a3 544#ifdef FIO_INC_DEBUG
4f1d43c2 545 fio_debug |= 1U << FD_TIME;
b5b571a3 546#endif
4f1d43c2 547 calibrate_cpu_clock();
b5b571a3 548#ifdef FIO_INC_DEBUG
4f1d43c2 549 fio_debug &= ~(1U << FD_TIME);
b5b571a3 550#endif
4f1d43c2 551
7d11f871
JA
552 threads = malloc(nr_cpus * sizeof(struct clock_thread));
553 tentries = CLOCK_ENTRIES * nr_cpus;
554 entries = malloc(tentries * sizeof(struct clock_entry));
555
556 log_info("cs: Testing %u CPUs\n", nr_cpus);
557
558 for (i = 0; i < nr_cpus; i++) {
559 struct clock_thread *t = &threads[i];
560
561 t->cpu = i;
562 t->seq = &seq;
563 t->entries = &entries[i * CLOCK_ENTRIES];
564 pthread_mutex_init(&t->lock, NULL);
565 pthread_mutex_init(&t->started, NULL);
566 pthread_mutex_lock(&t->lock);
6b0110cd
JA
567 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
568 failed++;
569 nr_cpus = i;
570 break;
571 }
7d11f871
JA
572 }
573
574 for (i = 0; i < nr_cpus; i++) {
575 struct clock_thread *t = &threads[i];
576
577 pthread_mutex_lock(&t->started);
578 }
579
580 for (i = 0; i < nr_cpus; i++) {
581 struct clock_thread *t = &threads[i];
582
583 pthread_mutex_unlock(&t->lock);
584 }
585
814917be 586 for (i = 0; i < nr_cpus; i++) {
7d11f871
JA
587 struct clock_thread *t = &threads[i];
588 void *ret;
589
590 pthread_join(t->thread, &ret);
591 if (ret)
592 failed++;
593 }
594 free(threads);
595
596 if (failed) {
4e0a8fa2 597 log_err("Clocksource test: %lu threads failed\n", failed);
7d11f871
JA
598 goto err;
599 }
600
601 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
602
603 for (failed = i = 0; i < tentries; i++) {
80da8a8f 604 this = &entries[i];
7d11f871
JA
605
606 if (!i) {
607 prev = this;
608 continue;
609 }
610
611 if (prev->tsc > this->tsc) {
612 uint64_t diff = prev->tsc - this->tsc;
613
4e0a8fa2
JA
614 log_info("cs: CPU clock mismatch (diff=%llu):\n",
615 (unsigned long long) diff);
616 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
617 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
7d11f871
JA
618 failed++;
619 }
620
621 prev = this;
622 }
623
624 if (failed)
625 log_info("cs: Failed: %lu\n", failed);
626 else
627 log_info("cs: Pass!\n");
628
629err:
630 free(entries);
631 return !!failed;
632}
633
634#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
635
636int fio_monotonic_clocktest(void)
637{
638 log_info("cs: current platform does not support CPU clocks\n");
639 return 0;
640}
641
642#endif