9988e6c54cd56a61200a17106f27a4512d9c6cb6
[fio.git] / gettime.c
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
18 static unsigned long cycles_per_usec;
19 static unsigned long inv_cycles_per_usec;
20 static uint64_t max_cycles_for_mult;
21 #endif
22 #ifdef ARCH_CPU_CLOCK_WRAPS
23 static unsigned long long cycles_start, cycles_wrap;
24 #endif
25 #endif
26 int tsc_reliable = 0;
27
28 struct tv_valid {
29         uint64_t last_cycles;
30         int last_tv_valid;
31         int warned;
32 };
33 #ifdef ARCH_HAVE_CPU_CLOCK
34 #ifdef CONFIG_TLS_THREAD
35 static __thread struct tv_valid static_tv_valid;
36 #else
37 static pthread_key_t tv_tls_key;
38 #endif
39 #endif
40
41 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
42 int fio_clock_source_set = 0;
43 static enum fio_cs fio_clock_source_inited = CS_INVAL;
44
45 #ifdef FIO_DEBUG_TIME
46
47 #define HASH_BITS       8
48 #define HASH_SIZE       (1 << HASH_BITS)
49
50 static struct flist_head hash[HASH_SIZE];
51 static int gtod_inited;
52
53 struct gtod_log {
54         struct flist_head list;
55         void *caller;
56         unsigned long calls;
57 };
58
59 static struct gtod_log *find_hash(void *caller)
60 {
61         unsigned long h = hash_ptr(caller, HASH_BITS);
62         struct flist_head *entry;
63
64         flist_for_each(entry, &hash[h]) {
65                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
66                                                                         list);
67
68                 if (log->caller == caller)
69                         return log;
70         }
71
72         return NULL;
73 }
74
75 static void inc_caller(void *caller)
76 {
77         struct gtod_log *log = find_hash(caller);
78
79         if (!log) {
80                 unsigned long h;
81
82                 log = malloc(sizeof(*log));
83                 INIT_FLIST_HEAD(&log->list);
84                 log->caller = caller;
85                 log->calls = 0;
86
87                 h = hash_ptr(caller, HASH_BITS);
88                 flist_add_tail(&log->list, &hash[h]);
89         }
90
91         log->calls++;
92 }
93
94 static void gtod_log_caller(void *caller)
95 {
96         if (gtod_inited)
97                 inc_caller(caller);
98 }
99
100 static void fio_exit fio_dump_gtod(void)
101 {
102         unsigned long total_calls = 0;
103         int i;
104
105         for (i = 0; i < HASH_SIZE; i++) {
106                 struct flist_head *entry;
107                 struct gtod_log *log;
108
109                 flist_for_each(entry, &hash[i]) {
110                         log = flist_entry(entry, struct gtod_log, list);
111
112                         printf("function %p, calls %lu\n", log->caller,
113                                                                 log->calls);
114                         total_calls += log->calls;
115                 }
116         }
117
118         printf("Total %lu gettimeofday\n", total_calls);
119 }
120
121 static void fio_init gtod_init(void)
122 {
123         int i;
124
125         for (i = 0; i < HASH_SIZE; i++)
126                 INIT_FLIST_HEAD(&hash[i]);
127
128         gtod_inited = 1;
129 }
130
131 #endif /* FIO_DEBUG_TIME */
132
133 #ifdef CONFIG_CLOCK_GETTIME
134 static int fill_clock_gettime(struct timespec *ts)
135 {
136 #if defined(CONFIG_CLOCK_MONOTONIC_RAW)
137         return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
138 #elif defined(CONFIG_CLOCK_MONOTONIC)
139         return clock_gettime(CLOCK_MONOTONIC, ts);
140 #else
141         return clock_gettime(CLOCK_REALTIME, ts);
142 #endif
143 }
144 #endif
145
146 static void __fio_gettime(struct timeval *tp)
147 {
148         switch (fio_clock_source) {
149 #ifdef CONFIG_GETTIMEOFDAY
150         case CS_GTOD:
151                 gettimeofday(tp, NULL);
152                 break;
153 #endif
154 #ifdef CONFIG_CLOCK_GETTIME
155         case CS_CGETTIME: {
156                 struct timespec ts;
157
158                 if (fill_clock_gettime(&ts) < 0) {
159                         log_err("fio: clock_gettime fails\n");
160                         assert(0);
161                 }
162
163                 tp->tv_sec = ts.tv_sec;
164                 tp->tv_usec = ts.tv_nsec / 1000;
165                 break;
166                 }
167 #endif
168 #ifdef ARCH_HAVE_CPU_CLOCK
169         case CS_CPUCLOCK: {
170                 uint64_t usecs, t;
171                 struct tv_valid *tv;
172
173 #ifdef CONFIG_TLS_THREAD
174                 tv = &static_tv_valid;
175 #else
176                 tv = pthread_getspecific(tv_tls_key);
177 #endif
178
179                 t = get_cpu_clock();
180 #ifdef ARCH_CPU_CLOCK_WRAPS
181                 if (t < cycles_start && !cycles_wrap)
182                         cycles_wrap = 1;
183                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
184                         log_err("fio: double CPU clock wrap\n");
185                         tv->warned = 1;
186                 }
187
188                 t -= cycles_start;
189 #endif
190                 tv->last_cycles = t;
191                 tv->last_tv_valid = 1;
192 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
193                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
194 #else
195                 if (t < max_cycles_for_mult)
196                         usecs = (t * inv_cycles_per_usec) / 16777216UL;
197                 else
198                         usecs = t / cycles_per_usec;
199 #endif
200                 tp->tv_sec = usecs / 1000000;
201                 tp->tv_usec = usecs % 1000000;
202                 break;
203                 }
204 #endif
205         default:
206                 log_err("fio: invalid clock source %d\n", fio_clock_source);
207                 break;
208         }
209 }
210
211 #ifdef FIO_DEBUG_TIME
212 void fio_gettime(struct timeval *tp, void *caller)
213 #else
214 void fio_gettime(struct timeval *tp, void fio_unused *caller)
215 #endif
216 {
217 #ifdef FIO_DEBUG_TIME
218         if (!caller)
219                 caller = __builtin_return_address(0);
220
221         gtod_log_caller(caller);
222 #endif
223         if (fio_unlikely(fio_gettime_offload(tp)))
224                 return;
225
226         __fio_gettime(tp);
227 }
228
229 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
230 static unsigned long get_cycles_per_usec(void)
231 {
232         struct timeval s, e;
233         uint64_t c_s, c_e;
234         enum fio_cs old_cs = fio_clock_source;
235
236 #ifdef CONFIG_CLOCK_GETTIME
237         fio_clock_source = CS_CGETTIME;
238 #else
239         fio_clock_source = CS_GTOD;
240 #endif
241         __fio_gettime(&s);
242
243         c_s = get_cpu_clock();
244         do {
245                 uint64_t elapsed;
246
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 + 127) >> 7;
258 }
259
260 #define NR_TIME_ITERS   50
261
262 static 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         mean /= 10.0;
304
305         for (i = 0; i < NR_TIME_ITERS; i++)
306                 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
307                                         (unsigned long long) cycles[i] / 10);
308
309         avg /= samples;
310         avg = (avg + 5) / 10;
311         minc /= 10;
312         maxc /= 10;
313         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
314         dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
315                         (unsigned long long) minc,
316                         (unsigned long long) maxc, mean, S);
317
318         cycles_per_usec = avg;
319         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
320         max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
321         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
322 #ifdef ARCH_CPU_CLOCK_WRAPS
323         cycles_start = get_cpu_clock();
324         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
325 #endif
326         return 0;
327 }
328 #else
329 static int calibrate_cpu_clock(void)
330 {
331 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
332         return 0;
333 #else
334         return 1;
335 #endif
336 }
337 #endif // ARCH_HAVE_CPU_CLOCK
338
339 #ifndef CONFIG_TLS_THREAD
340 void fio_local_clock_init(int is_thread)
341 {
342         struct tv_valid *t;
343
344         t = calloc(1, sizeof(*t));
345         if (pthread_setspecific(tv_tls_key, t)) {
346                 log_err("fio: can't set TLS key\n");
347                 assert(0);
348         }
349 }
350
351 static void kill_tv_tls_key(void *data)
352 {
353         free(data);
354 }
355 #else
356 void fio_local_clock_init(int is_thread)
357 {
358 }
359 #endif
360
361 void fio_clock_init(void)
362 {
363         if (fio_clock_source == fio_clock_source_inited)
364                 return;
365
366 #ifndef CONFIG_TLS_THREAD
367         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
368                 log_err("fio: can't create TLS key\n");
369 #endif
370
371         fio_clock_source_inited = fio_clock_source;
372
373         if (calibrate_cpu_clock())
374                 tsc_reliable = 0;
375
376         /*
377          * If the arch sets tsc_reliable != 0, then it must be good enough
378          * to use as THE clock source. For x86 CPUs, this means the TSC
379          * runs at a constant rate and is synced across CPU cores.
380          */
381         if (tsc_reliable) {
382                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
383                         fio_clock_source = CS_CPUCLOCK;
384         } else if (fio_clock_source == CS_CPUCLOCK)
385                 log_info("fio: clocksource=cpu may not be reliable\n");
386 }
387
388 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
389 {
390         long sec, usec;
391         uint64_t ret;
392
393         sec = e->tv_sec - s->tv_sec;
394         usec = e->tv_usec - s->tv_usec;
395         if (sec > 0 && usec < 0) {
396                 sec--;
397                 usec += 1000000;
398         }
399
400         /*
401          * time warp bug on some kernels?
402          */
403         if (sec < 0 || (sec == 0 && usec < 0))
404                 return 0;
405
406         ret = sec * 1000000ULL + usec;
407
408         return ret;
409 }
410
411 uint64_t utime_since_now(const struct timeval *s)
412 {
413         struct timeval t;
414
415         fio_gettime(&t, NULL);
416         return utime_since(s, &t);
417 }
418
419 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
420 {
421         long sec, usec, ret;
422
423         sec = e->tv_sec - s->tv_sec;
424         usec = e->tv_usec - s->tv_usec;
425         if (sec > 0 && usec < 0) {
426                 sec--;
427                 usec += 1000000;
428         }
429
430         if (sec < 0 || (sec == 0 && usec < 0))
431                 return 0;
432
433         sec *= 1000UL;
434         usec /= 1000UL;
435         ret = sec + usec;
436
437         return ret;
438 }
439
440 uint64_t mtime_since_now(const struct timeval *s)
441 {
442         struct timeval t;
443         void *p = __builtin_return_address(0);
444
445         fio_gettime(&t, p);
446         return mtime_since(s, &t);
447 }
448
449 uint64_t time_since_now(const struct timeval *s)
450 {
451         return mtime_since_now(s) / 1000;
452 }
453
454 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
455     defined(CONFIG_SFAA)
456
457 #define CLOCK_ENTRIES_DEBUG     100000
458 #define CLOCK_ENTRIES_TEST      10000
459
460 struct clock_entry {
461         uint32_t seq;
462         uint32_t cpu;
463         uint64_t tsc;
464 };
465
466 struct clock_thread {
467         pthread_t thread;
468         int cpu;
469         int debug;
470         pthread_mutex_t lock;
471         pthread_mutex_t started;
472         unsigned long nr_entries;
473         uint32_t *seq;
474         struct clock_entry *entries;
475 };
476
477 static inline uint32_t atomic32_inc_return(uint32_t *seq)
478 {
479         return 1 + __sync_fetch_and_add(seq, 1);
480 }
481
482 static void *clock_thread_fn(void *data)
483 {
484         struct clock_thread *t = data;
485         struct clock_entry *c;
486         os_cpu_mask_t cpu_mask;
487         uint32_t last_seq;
488         unsigned long long first;
489         int i;
490
491         if (fio_cpuset_init(&cpu_mask)) {
492                 int __err = errno;
493
494                 log_err("clock cpuset init failed: %s\n", strerror(__err));
495                 goto err_out;
496         }
497
498         fio_cpu_set(&cpu_mask, t->cpu);
499
500         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
501                 int __err = errno;
502
503                 log_err("clock setaffinity failed: %s\n", strerror(__err));
504                 goto err;
505         }
506
507         pthread_mutex_lock(&t->lock);
508         pthread_mutex_unlock(&t->started);
509
510         first = get_cpu_clock();
511         last_seq = 0;
512         c = &t->entries[0];
513         for (i = 0; i < t->nr_entries; i++, c++) {
514                 uint32_t seq;
515                 uint64_t tsc;
516
517                 c->cpu = t->cpu;
518                 do {
519                         seq = atomic32_inc_return(t->seq);
520                         if (seq < last_seq)
521                                 break;
522                         tsc = get_cpu_clock();
523                 } while (seq != *t->seq);
524
525                 c->seq = seq;
526                 c->tsc = tsc;
527         }
528
529         if (t->debug) {
530                 unsigned long long clocks;
531
532                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
533                 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
534                                                         clocks, first);
535         }
536
537         /*
538          * The most common platform clock breakage is returning zero
539          * indefinitely. Check for that and return failure.
540          */
541         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
542                 goto err;
543
544         fio_cpuset_exit(&cpu_mask);
545         return NULL;
546 err:
547         fio_cpuset_exit(&cpu_mask);
548 err_out:
549         return (void *) 1;
550 }
551
552 static int clock_cmp(const void *p1, const void *p2)
553 {
554         const struct clock_entry *c1 = p1;
555         const struct clock_entry *c2 = p2;
556
557         if (c1->seq == c2->seq)
558                 log_err("cs: bug in atomic sequence!\n");
559
560         return c1->seq - c2->seq;
561 }
562
563 int fio_monotonic_clocktest(int debug)
564 {
565         struct clock_thread *cthreads;
566         unsigned int nr_cpus = cpus_online();
567         struct clock_entry *entries;
568         unsigned long nr_entries, tentries, failed = 0;
569         struct clock_entry *prev, *this;
570         uint32_t seq = 0;
571         unsigned int i;
572
573         if (debug) {
574                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
575
576 #ifdef FIO_INC_DEBUG
577                 fio_debug |= 1U << FD_TIME;
578 #endif
579                 nr_entries = CLOCK_ENTRIES_DEBUG;
580         } else
581                 nr_entries = CLOCK_ENTRIES_TEST;
582
583         calibrate_cpu_clock();
584
585         if (debug) {
586 #ifdef FIO_INC_DEBUG
587                 fio_debug &= ~(1U << FD_TIME);
588 #endif
589         }
590
591         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
592         tentries = nr_entries * nr_cpus;
593         entries = malloc(tentries * sizeof(struct clock_entry));
594
595         if (debug)
596                 log_info("cs: Testing %u CPUs\n", nr_cpus);
597
598         for (i = 0; i < nr_cpus; i++) {
599                 struct clock_thread *t = &cthreads[i];
600
601                 t->cpu = i;
602                 t->debug = debug;
603                 t->seq = &seq;
604                 t->nr_entries = nr_entries;
605                 t->entries = &entries[i * nr_entries];
606                 pthread_mutex_init(&t->lock, NULL);
607                 pthread_mutex_init(&t->started, NULL);
608                 pthread_mutex_lock(&t->lock);
609                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
610                         failed++;
611                         nr_cpus = i;
612                         break;
613                 }
614         }
615
616         for (i = 0; i < nr_cpus; i++) {
617                 struct clock_thread *t = &cthreads[i];
618
619                 pthread_mutex_lock(&t->started);
620         }
621
622         for (i = 0; i < nr_cpus; i++) {
623                 struct clock_thread *t = &cthreads[i];
624
625                 pthread_mutex_unlock(&t->lock);
626         }
627
628         for (i = 0; i < nr_cpus; i++) {
629                 struct clock_thread *t = &cthreads[i];
630                 void *ret;
631
632                 pthread_join(t->thread, &ret);
633                 if (ret)
634                         failed++;
635         }
636         free(cthreads);
637
638         if (failed) {
639                 if (debug)
640                         log_err("Clocksource test: %lu threads failed\n", failed);
641                 goto err;
642         }
643
644         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
645
646         /* silence silly gcc */
647         prev = NULL;
648         for (failed = i = 0; i < tentries; i++) {
649                 this = &entries[i];
650
651                 if (!i) {
652                         prev = this;
653                         continue;
654                 }
655
656                 if (prev->tsc > this->tsc) {
657                         uint64_t diff = prev->tsc - this->tsc;
658
659                         if (!debug) {
660                                 failed++;
661                                 break;
662                         }
663
664                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
665                                                 (unsigned long long) diff);
666                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
667                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
668                         failed++;
669                 }
670
671                 prev = this;
672         }
673
674         if (debug) {
675                 if (failed)
676                         log_info("cs: Failed: %lu\n", failed);
677                 else
678                         log_info("cs: Pass!\n");
679         }
680 err:
681         free(entries);
682         return !!failed;
683 }
684
685 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
686
687 int fio_monotonic_clocktest(int debug)
688 {
689         if (debug)
690                 log_info("cs: current platform does not support CPU clocks\n");
691         return 1;
692 }
693
694 #endif