Fix problem with terminating on unaligned sizes
[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 #ifdef ARCH_HAVE_CPU_CLOCK
17 static unsigned long cycles_per_usec;
18 static unsigned long inv_cycles_per_usec;
19 #endif
20 int tsc_reliable = 0;
21
22 struct tv_valid {
23         struct timeval last_tv;
24         uint64_t last_cycles;
25         int last_tv_valid;
26 };
27 #ifdef CONFIG_TLS_THREAD
28 static struct tv_valid __thread static_tv_valid;
29 #else
30 static pthread_key_t tv_tls_key;
31 #endif
32
33 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
34 int fio_clock_source_set = 0;
35 enum fio_cs fio_clock_source_inited = CS_INVAL;
36
37 #ifdef FIO_DEBUG_TIME
38
39 #define HASH_BITS       8
40 #define HASH_SIZE       (1 << HASH_BITS)
41
42 static struct flist_head hash[HASH_SIZE];
43 static int gtod_inited;
44
45 struct gtod_log {
46         struct flist_head list;
47         void *caller;
48         unsigned long calls;
49 };
50
51 static struct gtod_log *find_hash(void *caller)
52 {
53         unsigned long h = hash_ptr(caller, HASH_BITS);
54         struct flist_head *entry;
55
56         flist_for_each(entry, &hash[h]) {
57                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
58                                                                         list);
59
60                 if (log->caller == caller)
61                         return log;
62         }
63
64         return NULL;
65 }
66
67 static 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));
75                 INIT_FLIST_HEAD(&log->list);
76                 log->caller = caller;
77                 log->calls = 0;
78
79                 h = hash_ptr(caller, HASH_BITS);
80                 flist_add_tail(&log->list, &hash[h]);
81         }
82
83         return log;
84 }
85
86 static 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
95 static 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++) {
101                 struct flist_head *entry;
102                 struct gtod_log *log;
103
104                 flist_for_each(entry, &hash[i]) {
105                         log = flist_entry(entry, struct gtod_log, list);
106
107                         printf("function %p, calls %lu\n", log->caller,
108                                                                 log->calls);
109                         total_calls += log->calls;
110                 }
111         }
112
113         printf("Total %lu gettimeofday\n", total_calls);
114 }
115
116 static void fio_init gtod_init(void)
117 {
118         int i;
119
120         for (i = 0; i < HASH_SIZE; i++)
121                 INIT_FLIST_HEAD(&hash[i]);
122
123         gtod_inited = 1;
124 }
125
126 #endif /* FIO_DEBUG_TIME */
127
128 #ifdef CONFIG_CLOCK_GETTIME
129 static int fill_clock_gettime(struct timespec *ts)
130 {
131 #ifdef CONFIG_CLOCK_MONOTONIC
132         return clock_gettime(CLOCK_MONOTONIC, ts);
133 #else
134         return clock_gettime(CLOCK_REALTIME, ts);
135 #endif
136 }
137 #endif
138
139 static void *__fio_gettime(struct timeval *tp)
140 {
141         struct tv_valid *tv;
142
143 #ifdef CONFIG_TLS_THREAD
144         tv = &static_tv_valid;
145 #else
146         tv = pthread_getspecific(tv_tls_key);
147 #endif
148
149         switch (fio_clock_source) {
150 #ifdef CONFIG_GETTIMEOFDAY
151         case CS_GTOD:
152                 gettimeofday(tp, NULL);
153                 break;
154 #endif
155 #ifdef CONFIG_CLOCK_GETTIME
156         case CS_CGETTIME: {
157                 struct timespec ts;
158
159                 if (fill_clock_gettime(&ts) < 0) {
160                         log_err("fio: clock_gettime fails\n");
161                         assert(0);
162                 }
163
164                 tp->tv_sec = ts.tv_sec;
165                 tp->tv_usec = ts.tv_nsec / 1000;
166                 break;
167                 }
168 #endif
169 #ifdef ARCH_HAVE_CPU_CLOCK
170         case CS_CPUCLOCK: {
171                 uint64_t usecs, t;
172
173                 t = get_cpu_clock();
174                 if (tv && t < tv->last_cycles) {
175                         dprint(FD_TIME, "CPU clock going back in time\n");
176                         t = tv->last_cycles;
177                 } else if (tv)
178                         tv->last_cycles = t;
179
180                 usecs = (t * inv_cycles_per_usec) / 16777216UL;
181                 tp->tv_sec = usecs / 1000000;
182                 tp->tv_usec = usecs % 1000000;
183                 break;
184                 }
185 #endif
186         default:
187                 log_err("fio: invalid clock source %d\n", fio_clock_source);
188                 break;
189         }
190
191         return tv;
192 }
193
194 #ifdef FIO_DEBUG_TIME
195 void fio_gettime(struct timeval *tp, void *caller)
196 #else
197 void fio_gettime(struct timeval *tp, void fio_unused *caller)
198 #endif
199 {
200         struct tv_valid *tv;
201
202 #ifdef FIO_DEBUG_TIME
203         if (!caller)
204                 caller = __builtin_return_address(0);
205
206         gtod_log_caller(caller);
207 #endif
208         if (fio_tv) {
209                 memcpy(tp, fio_tv, sizeof(*tp));
210                 return;
211         }
212
213         tv = __fio_gettime(tp);
214
215         /*
216          * If Linux is using the tsc clock on non-synced processors,
217          * sometimes time can appear to drift backwards. Fix that up.
218          */
219         if (tv) {
220                 if (tv->last_tv_valid) {
221                         if (tp->tv_sec < tv->last_tv.tv_sec)
222                                 tp->tv_sec = tv->last_tv.tv_sec;
223                         else if (tv->last_tv.tv_sec == tp->tv_sec &&
224                                  tp->tv_usec < tv->last_tv.tv_usec)
225                                 tp->tv_usec = tv->last_tv.tv_usec;
226                 }
227                 tv->last_tv_valid = 1;
228                 memcpy(&tv->last_tv, tp, sizeof(*tp));
229         }
230 }
231
232 #ifdef ARCH_HAVE_CPU_CLOCK
233 static unsigned long get_cycles_per_usec(void)
234 {
235         struct timeval s, e;
236         uint64_t c_s, c_e;
237         enum fio_cs old_cs = fio_clock_source;
238
239 #ifdef CONFIG_CLOCK_GETTIME
240         fio_clock_source = CS_CGETTIME;
241 #else
242         fio_clock_source = CS_GTOD;
243 #endif
244         __fio_gettime(&s);
245
246         c_s = get_cpu_clock();
247         do {
248                 uint64_t elapsed;
249
250                 __fio_gettime(&e);
251
252                 elapsed = utime_since(&s, &e);
253                 if (elapsed >= 1280) {
254                         c_e = get_cpu_clock();
255                         break;
256                 }
257         } while (1);
258
259         fio_clock_source = old_cs;
260         return (c_e - c_s + 127) >> 7;
261 }
262
263 #define NR_TIME_ITERS   50
264
265 static void calibrate_cpu_clock(void)
266 {
267         double delta, mean, S;
268         uint64_t avg, cycles[NR_TIME_ITERS];
269         int i, samples;
270
271         cycles[0] = get_cycles_per_usec();
272         S = delta = mean = 0.0;
273         for (i = 0; i < NR_TIME_ITERS; i++) {
274                 cycles[i] = get_cycles_per_usec();
275                 delta = cycles[i] - mean;
276                 if (delta) {
277                         mean += delta / (i + 1.0);
278                         S += delta * (cycles[i] - mean);
279                 }
280         }
281
282         S = sqrt(S / (NR_TIME_ITERS - 1.0));
283
284         samples = avg = 0;
285         for (i = 0; i < NR_TIME_ITERS; i++) {
286                 double this = cycles[i];
287
288                 if ((fmax(this, mean) - fmin(this, mean)) > S)
289                         continue;
290                 samples++;
291                 avg += this;
292         }
293
294         S /= (double) NR_TIME_ITERS;
295         mean /= 10.0;
296
297         for (i = 0; i < NR_TIME_ITERS; i++)
298                 dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
299
300         avg /= samples;
301         avg = (avg + 5) / 10;
302         dprint(FD_TIME, "avg: %lu\n", avg);
303         dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
304
305         cycles_per_usec = avg;
306         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
307         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
308 }
309 #else
310 static void calibrate_cpu_clock(void)
311 {
312 }
313 #endif
314
315 #ifndef CONFIG_TLS_THREAD
316 void fio_local_clock_init(int is_thread)
317 {
318         struct tv_valid *t;
319
320         t = calloc(sizeof(*t), 1);
321         if (pthread_setspecific(tv_tls_key, t))
322                 log_err("fio: can't set TLS key\n");
323 }
324
325 static void kill_tv_tls_key(void *data)
326 {
327         free(data);
328 }
329 #else
330 void fio_local_clock_init(int is_thread)
331 {
332 }
333 #endif
334
335 void fio_clock_init(void)
336 {
337         if (fio_clock_source == fio_clock_source_inited)
338                 return;
339
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");
343 #endif
344
345         fio_clock_source_inited = fio_clock_source;
346         calibrate_cpu_clock();
347
348         /*
349          * If the arch sets tsc_reliable != 0, then it must be good enough
350          * to use as THE clock source. For x86 CPUs, this means the TSC
351          * runs at a constant rate and is synced across CPU cores.
352          */
353         if (tsc_reliable) {
354                 if (!fio_clock_source_set)
355                         fio_clock_source = CS_CPUCLOCK;
356         } else if (fio_clock_source == CS_CPUCLOCK)
357                 log_info("fio: clocksource=cpu may not be reliable\n");
358 }
359
360 uint64_t utime_since(struct timeval *s, struct timeval *e)
361 {
362         long sec, usec;
363         uint64_t ret;
364
365         sec = e->tv_sec - s->tv_sec;
366         usec = e->tv_usec - s->tv_usec;
367         if (sec > 0 && usec < 0) {
368                 sec--;
369                 usec += 1000000;
370         }
371
372         /*
373          * time warp bug on some kernels?
374          */
375         if (sec < 0 || (sec == 0 && usec < 0))
376                 return 0;
377
378         ret = sec * 1000000ULL + usec;
379
380         return ret;
381 }
382
383 uint64_t utime_since_now(struct timeval *s)
384 {
385         struct timeval t;
386
387         fio_gettime(&t, NULL);
388         return utime_since(s, &t);
389 }
390
391 uint64_t mtime_since(struct timeval *s, struct timeval *e)
392 {
393         long sec, usec, ret;
394
395         sec = e->tv_sec - s->tv_sec;
396         usec = e->tv_usec - s->tv_usec;
397         if (sec > 0 && usec < 0) {
398                 sec--;
399                 usec += 1000000;
400         }
401
402         if (sec < 0 || (sec == 0 && usec < 0))
403                 return 0;
404
405         sec *= 1000UL;
406         usec /= 1000UL;
407         ret = sec + usec;
408
409         return ret;
410 }
411
412 uint64_t mtime_since_now(struct timeval *s)
413 {
414         struct timeval t;
415         void *p = __builtin_return_address(0);
416
417         fio_gettime(&t, p);
418         return mtime_since(s, &t);
419 }
420
421 uint64_t time_since_now(struct timeval *s)
422 {
423         return mtime_since_now(s) / 1000;
424 }
425
426 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
427     defined(CONFIG_SFAA)
428
429 #define CLOCK_ENTRIES   100000
430
431 struct clock_entry {
432         uint64_t seq;
433         uint64_t tsc;
434         uint64_t cpu;
435 };
436
437 struct clock_thread {
438         pthread_t thread;
439         int cpu;
440         pthread_mutex_t lock;
441         pthread_mutex_t started;
442         uint64_t *seq;
443         struct clock_entry *entries;
444 };
445
446 static inline uint64_t atomic64_inc_return(uint64_t *seq)
447 {
448         return 1 + __sync_fetch_and_add(seq, 1);
449 }
450
451 static void *clock_thread_fn(void *data)
452 {
453         struct clock_thread *t = data;
454         struct clock_entry *c;
455         os_cpu_mask_t cpu_mask;
456         int i;
457
458         memset(&cpu_mask, 0, sizeof(cpu_mask));
459         fio_cpu_set(&cpu_mask, t->cpu);
460
461         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
462                 log_err("clock setaffinity failed\n");
463                 return (void *) 1;
464         }
465
466         pthread_mutex_lock(&t->lock);
467         pthread_mutex_unlock(&t->started);
468
469         c = &t->entries[0];
470         for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
471                 uint64_t seq, tsc;
472
473                 c->cpu = t->cpu;
474                 do {
475                         seq = atomic64_inc_return(t->seq);
476                         tsc = get_cpu_clock();
477                 } while (seq != *t->seq);
478
479                 c->seq = seq;
480                 c->tsc = tsc;
481         }
482
483         log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
484         return NULL;
485 }
486
487 static int clock_cmp(const void *p1, const void *p2)
488 {
489         const struct clock_entry *c1 = p1;
490         const struct clock_entry *c2 = p2;
491
492         if (c1->seq == c2->seq)
493                 log_err("cs: bug in atomic sequence!\n");
494
495         return c1->seq - c2->seq;
496 }
497
498 int fio_monotonic_clocktest(void)
499 {
500         struct clock_thread *threads;
501         unsigned int nr_cpus = cpus_online();
502         struct clock_entry *entries;
503         unsigned long tentries, failed;
504         struct clock_entry *prev, *this;
505         uint64_t seq = 0;
506         int i;
507
508         log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
509
510         fio_debug |= 1U << FD_TIME;
511         calibrate_cpu_clock();
512         fio_debug &= ~(1U << FD_TIME);
513
514         threads = malloc(nr_cpus * sizeof(struct clock_thread));
515         tentries = CLOCK_ENTRIES * nr_cpus;
516         entries = malloc(tentries * sizeof(struct clock_entry));
517
518         log_info("cs: Testing %u CPUs\n", nr_cpus);
519
520         for (i = 0; i < nr_cpus; i++) {
521                 struct clock_thread *t = &threads[i];
522
523                 t->cpu = i;
524                 t->seq = &seq;
525                 t->entries = &entries[i * CLOCK_ENTRIES];
526                 pthread_mutex_init(&t->lock, NULL);
527                 pthread_mutex_init(&t->started, NULL);
528                 pthread_mutex_lock(&t->lock);
529                 pthread_create(&t->thread, NULL, clock_thread_fn, t);
530         }
531
532         for (i = 0; i < nr_cpus; i++) {
533                 struct clock_thread *t = &threads[i];
534
535                 pthread_mutex_lock(&t->started);
536         }
537
538         for (i = 0; i < nr_cpus; i++) {
539                 struct clock_thread *t = &threads[i];
540
541                 pthread_mutex_unlock(&t->lock);
542         }
543
544         for (failed = i = 0; i < nr_cpus; i++) {
545                 struct clock_thread *t = &threads[i];
546                 void *ret;
547
548                 pthread_join(t->thread, &ret);
549                 if (ret)
550                         failed++;
551         }
552         free(threads);
553
554         if (failed) {
555                 log_err("Clocksource test: %u threads failed\n", failed);
556                 goto err;
557         }
558
559         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
560
561         for (failed = i = 0; i < tentries; i++) {
562                 this = &entries[i];
563
564                 if (!i) {
565                         prev = this;
566                         continue;
567                 }
568
569                 if (prev->tsc > this->tsc) {
570                         uint64_t diff = prev->tsc - this->tsc;
571
572                         log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
573                         log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
574                         log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
575                         failed++;
576                 }
577
578                 prev = this;
579         }
580
581         if (failed)
582                 log_info("cs: Failed: %lu\n", failed);
583         else
584                 log_info("cs: Pass!\n");
585
586 err:
587         free(entries);
588         return !!failed;
589 }
590
591 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
592
593 int fio_monotonic_clocktest(void)
594 {
595         log_info("cs: current platform does not support CPU clocks\n");
596         return 0;
597 }
598
599 #endif