Commit | Line | Data |
---|---|---|
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 | |
09a32402 | 16 | #ifdef ARCH_HAVE_CPU_CLOCK |
c223da83 | 17 | static unsigned long cycles_per_usec; |
c223da83 | 18 | static unsigned long last_cycles; |
fa80feae | 19 | int tsc_reliable = 0; |
09a32402 JA |
20 | #endif |
21 | static struct timeval last_tv; | |
3e488920 | 22 | static int last_tv_valid; |
02bcaa8c | 23 | |
16de1bf9 | 24 | enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE; |
fa80feae | 25 | int fio_clock_source_set = 0; |
01423eae | 26 | enum fio_cs fio_clock_source_inited = CS_INVAL; |
c223da83 | 27 | |
02bcaa8c JA |
28 | #ifdef FIO_DEBUG_TIME |
29 | ||
30 | #define HASH_BITS 8 | |
31 | #define HASH_SIZE (1 << HASH_BITS) | |
32 | ||
01743ee1 | 33 | static struct flist_head hash[HASH_SIZE]; |
02bcaa8c JA |
34 | static int gtod_inited; |
35 | ||
36 | struct gtod_log { | |
01743ee1 | 37 | struct flist_head list; |
02bcaa8c JA |
38 | void *caller; |
39 | unsigned long calls; | |
40 | }; | |
41 | ||
42 | static struct gtod_log *find_hash(void *caller) | |
43 | { | |
44 | unsigned long h = hash_ptr(caller, HASH_BITS); | |
01743ee1 | 45 | struct flist_head *entry; |
02bcaa8c | 46 | |
01743ee1 JA |
47 | flist_for_each(entry, &hash[h]) { |
48 | struct gtod_log *log = flist_entry(entry, struct gtod_log, | |
49 | list); | |
02bcaa8c JA |
50 | |
51 | if (log->caller == caller) | |
52 | return log; | |
53 | } | |
54 | ||
55 | return NULL; | |
56 | } | |
57 | ||
58 | static struct gtod_log *find_log(void *caller) | |
59 | { | |
60 | struct gtod_log *log = find_hash(caller); | |
61 | ||
62 | if (!log) { | |
63 | unsigned long h; | |
64 | ||
65 | log = malloc(sizeof(*log)); | |
01743ee1 | 66 | INIT_FLIST_HEAD(&log->list); |
02bcaa8c JA |
67 | log->caller = caller; |
68 | log->calls = 0; | |
69 | ||
70 | h = hash_ptr(caller, HASH_BITS); | |
01743ee1 | 71 | flist_add_tail(&log->list, &hash[h]); |
02bcaa8c JA |
72 | } |
73 | ||
74 | return log; | |
75 | } | |
76 | ||
77 | static void gtod_log_caller(void *caller) | |
78 | { | |
79 | if (gtod_inited) { | |
80 | struct gtod_log *log = find_log(caller); | |
81 | ||
82 | log->calls++; | |
83 | } | |
84 | } | |
85 | ||
86 | static void fio_exit fio_dump_gtod(void) | |
87 | { | |
88 | unsigned long total_calls = 0; | |
89 | int i; | |
90 | ||
91 | for (i = 0; i < HASH_SIZE; i++) { | |
01743ee1 | 92 | struct flist_head *entry; |
02bcaa8c JA |
93 | struct gtod_log *log; |
94 | ||
01743ee1 JA |
95 | flist_for_each(entry, &hash[i]) { |
96 | log = flist_entry(entry, struct gtod_log, list); | |
02bcaa8c | 97 | |
5ec10eaa JA |
98 | printf("function %p, calls %lu\n", log->caller, |
99 | log->calls); | |
02bcaa8c JA |
100 | total_calls += log->calls; |
101 | } | |
102 | } | |
103 | ||
104 | printf("Total %lu gettimeofday\n", total_calls); | |
105 | } | |
106 | ||
107 | static void fio_init gtod_init(void) | |
108 | { | |
109 | int i; | |
110 | ||
111 | for (i = 0; i < HASH_SIZE; i++) | |
01743ee1 | 112 | INIT_FLIST_HEAD(&hash[i]); |
02bcaa8c JA |
113 | |
114 | gtod_inited = 1; | |
115 | } | |
116 | ||
117 | #endif /* FIO_DEBUG_TIME */ | |
118 | ||
1e97cce9 | 119 | #ifdef FIO_DEBUG_TIME |
02bcaa8c | 120 | void fio_gettime(struct timeval *tp, void *caller) |
1e97cce9 JA |
121 | #else |
122 | void fio_gettime(struct timeval *tp, void fio_unused *caller) | |
123 | #endif | |
02bcaa8c JA |
124 | { |
125 | #ifdef FIO_DEBUG_TIME | |
126 | if (!caller) | |
127 | caller = __builtin_return_address(0); | |
128 | ||
129 | gtod_log_caller(caller); | |
02bcaa8c | 130 | #endif |
be4ecfdf JA |
131 | if (fio_tv) { |
132 | memcpy(tp, fio_tv, sizeof(*tp)); | |
133 | return; | |
c223da83 JA |
134 | } |
135 | ||
136 | switch (fio_clock_source) { | |
137 | case CS_GTOD: | |
02bcaa8c | 138 | gettimeofday(tp, NULL); |
c223da83 JA |
139 | break; |
140 | case CS_CGETTIME: { | |
02bcaa8c JA |
141 | struct timespec ts; |
142 | ||
5bd9c78c JA |
143 | #ifdef FIO_HAVE_CLOCK_MONOTONIC |
144 | if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) { | |
145 | #else | |
d481e006 | 146 | if (clock_gettime(CLOCK_REALTIME, &ts) < 0) { |
5bd9c78c | 147 | #endif |
c223da83 JA |
148 | log_err("fio: clock_gettime fails\n"); |
149 | assert(0); | |
02bcaa8c JA |
150 | } |
151 | ||
152 | tp->tv_sec = ts.tv_sec; | |
153 | tp->tv_usec = ts.tv_nsec / 1000; | |
c223da83 JA |
154 | break; |
155 | } | |
156 | #ifdef ARCH_HAVE_CPU_CLOCK | |
157 | case CS_CPUCLOCK: { | |
158 | unsigned long long usecs, t; | |
159 | ||
160 | t = get_cpu_clock(); | |
161 | if (t < last_cycles) { | |
162 | dprint(FD_TIME, "CPU clock going back in time\n"); | |
163 | t = last_cycles; | |
164 | } | |
165 | ||
166 | usecs = t / cycles_per_usec; | |
167 | tp->tv_sec = usecs / 1000000; | |
168 | tp->tv_usec = usecs % 1000000; | |
169 | last_cycles = t; | |
170 | break; | |
171 | } | |
172 | #endif | |
173 | default: | |
174 | log_err("fio: invalid clock source %d\n", fio_clock_source); | |
175 | break; | |
02bcaa8c | 176 | } |
3e488920 JA |
177 | |
178 | /* | |
179 | * If Linux is using the tsc clock on non-synced processors, | |
180 | * sometimes time can appear to drift backwards. Fix that up. | |
181 | */ | |
182 | if (last_tv_valid) { | |
183 | if (tp->tv_sec < last_tv.tv_sec) | |
184 | tp->tv_sec = last_tv.tv_sec; | |
185 | else if (last_tv.tv_sec == tp->tv_sec && | |
186 | tp->tv_usec < last_tv.tv_usec) | |
187 | tp->tv_usec = last_tv.tv_usec; | |
188 | } | |
189 | last_tv_valid = 1; | |
190 | memcpy(&last_tv, tp, sizeof(*tp)); | |
02bcaa8c | 191 | } |
be4ecfdf | 192 | |
09a32402 | 193 | #ifdef ARCH_HAVE_CPU_CLOCK |
c223da83 JA |
194 | static unsigned long get_cycles_per_usec(void) |
195 | { | |
196 | struct timeval s, e; | |
197 | unsigned long long c_s, c_e; | |
198 | ||
199 | gettimeofday(&s, NULL); | |
200 | c_s = get_cpu_clock(); | |
201 | do { | |
202 | unsigned long long elapsed; | |
203 | ||
204 | gettimeofday(&e, NULL); | |
205 | elapsed = utime_since(&s, &e); | |
486332e5 | 206 | if (elapsed >= 1280) { |
c223da83 JA |
207 | c_e = get_cpu_clock(); |
208 | break; | |
209 | } | |
210 | } while (1); | |
211 | ||
c011000d | 212 | return (c_e - c_s + 127) >> 7; |
c223da83 JA |
213 | } |
214 | ||
fa80feae JA |
215 | #define NR_TIME_ITERS 50 |
216 | ||
09a32402 | 217 | static void calibrate_cpu_clock(void) |
c223da83 JA |
218 | { |
219 | double delta, mean, S; | |
fa80feae | 220 | unsigned long avg, cycles[NR_TIME_ITERS]; |
c223da83 JA |
221 | int i, samples; |
222 | ||
c223da83 JA |
223 | cycles[0] = get_cycles_per_usec(); |
224 | S = delta = mean = 0.0; | |
fa80feae | 225 | for (i = 0; i < NR_TIME_ITERS; i++) { |
c223da83 JA |
226 | cycles[i] = get_cycles_per_usec(); |
227 | delta = cycles[i] - mean; | |
228 | if (delta) { | |
229 | mean += delta / (i + 1.0); | |
230 | S += delta * (cycles[i] - mean); | |
231 | } | |
232 | } | |
233 | ||
fa80feae | 234 | S = sqrt(S / (NR_TIME_ITERS - 1.0)); |
c223da83 JA |
235 | |
236 | samples = avg = 0; | |
fa80feae | 237 | for (i = 0; i < NR_TIME_ITERS; i++) { |
c223da83 JA |
238 | double this = cycles[i]; |
239 | ||
03e20d68 | 240 | if ((fmax(this, mean) - fmin(this, mean)) > S) |
c223da83 JA |
241 | continue; |
242 | samples++; | |
243 | avg += this; | |
244 | } | |
245 | ||
fa80feae | 246 | S /= (double) NR_TIME_ITERS; |
89db727d | 247 | mean /= 10.0; |
c223da83 | 248 | |
fa80feae | 249 | for (i = 0; i < NR_TIME_ITERS; i++) |
c223da83 JA |
250 | dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10); |
251 | ||
d7abad3d JA |
252 | avg /= samples; |
253 | avg = (avg + 9) / 10; | |
c223da83 JA |
254 | dprint(FD_TIME, "avg: %lu\n", avg); |
255 | dprint(FD_TIME, "mean=%f, S=%f\n", mean, S); | |
256 | ||
257 | cycles_per_usec = avg; | |
09a32402 JA |
258 | } |
259 | #else | |
260 | static void calibrate_cpu_clock(void) | |
261 | { | |
262 | } | |
263 | #endif | |
264 | ||
265 | void fio_clock_init(void) | |
266 | { | |
01423eae JA |
267 | if (fio_clock_source == fio_clock_source_inited) |
268 | return; | |
269 | ||
09a32402 | 270 | last_tv_valid = 0; |
01423eae | 271 | fio_clock_source_inited = fio_clock_source; |
09a32402 | 272 | calibrate_cpu_clock(); |
fa80feae JA |
273 | |
274 | /* | |
275 | * If the arch sets tsc_reliable != 0, then it must be good enough | |
276 | * to use as THE clock source. For x86 CPUs, this means the TSC | |
277 | * runs at a constant rate and is synced across CPU cores. | |
278 | */ | |
279 | if (tsc_reliable) { | |
280 | if (!fio_clock_source_set) | |
281 | fio_clock_source = CS_CPUCLOCK; | |
282 | } else if (fio_clock_source == CS_CPUCLOCK) | |
283 | log_info("fio: clocksource=cpu may not be reliable\n"); | |
c223da83 JA |
284 | } |
285 | ||
39ab7da2 | 286 | unsigned long long utime_since(struct timeval *s, struct timeval *e) |
be4ecfdf | 287 | { |
39ab7da2 JA |
288 | long sec, usec; |
289 | unsigned long long ret; | |
290 | ||
291 | sec = e->tv_sec - s->tv_sec; | |
292 | usec = e->tv_usec - s->tv_usec; | |
293 | if (sec > 0 && usec < 0) { | |
294 | sec--; | |
295 | usec += 1000000; | |
296 | } | |
297 | ||
298 | /* | |
299 | * time warp bug on some kernels? | |
300 | */ | |
301 | if (sec < 0 || (sec == 0 && usec < 0)) | |
302 | return 0; | |
303 | ||
304 | ret = sec * 1000000ULL + usec; | |
305 | ||
306 | return ret; | |
be4ecfdf JA |
307 | } |
308 | ||
39ab7da2 | 309 | unsigned long long utime_since_now(struct timeval *s) |
be4ecfdf | 310 | { |
39ab7da2 JA |
311 | struct timeval t; |
312 | ||
313 | fio_gettime(&t, NULL); | |
314 | return utime_since(s, &t); | |
be4ecfdf | 315 | } |
783a3eb1 | 316 | |
39ab7da2 | 317 | unsigned long mtime_since(struct timeval *s, struct timeval *e) |
783a3eb1 | 318 | { |
39ab7da2 | 319 | long sec, usec, ret; |
783a3eb1 | 320 | |
39ab7da2 JA |
321 | sec = e->tv_sec - s->tv_sec; |
322 | usec = e->tv_usec - s->tv_usec; | |
323 | if (sec > 0 && usec < 0) { | |
324 | sec--; | |
325 | usec += 1000000; | |
783a3eb1 JA |
326 | } |
327 | ||
39ab7da2 JA |
328 | if (sec < 0 || (sec == 0 && usec < 0)) |
329 | return 0; | |
330 | ||
331 | sec *= 1000UL; | |
332 | usec /= 1000UL; | |
333 | ret = sec + usec; | |
334 | ||
335 | return ret; | |
783a3eb1 JA |
336 | } |
337 | ||
39ab7da2 | 338 | unsigned long mtime_since_now(struct timeval *s) |
783a3eb1 | 339 | { |
39ab7da2 JA |
340 | struct timeval t; |
341 | void *p = __builtin_return_address(0); | |
783a3eb1 | 342 | |
39ab7da2 JA |
343 | fio_gettime(&t, p); |
344 | return mtime_since(s, &t); | |
345 | } | |
783a3eb1 | 346 | |
39ab7da2 JA |
347 | unsigned long time_since_now(struct timeval *s) |
348 | { | |
349 | return mtime_since_now(s) / 1000; | |
783a3eb1 | 350 | } |
7d11f871 JA |
351 | |
352 | #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) | |
353 | ||
354 | #define CLOCK_ENTRIES 100000 | |
355 | ||
356 | struct clock_entry { | |
357 | unsigned long seq; | |
358 | unsigned long tsc; | |
359 | unsigned long cpu; | |
360 | }; | |
361 | ||
362 | struct clock_thread { | |
363 | pthread_t thread; | |
364 | int cpu; | |
365 | pthread_mutex_t lock; | |
366 | pthread_mutex_t started; | |
367 | uint64_t *seq; | |
368 | struct clock_entry *entries; | |
369 | }; | |
370 | ||
371 | static inline uint64_t atomic64_inc_return(uint64_t *seq) | |
372 | { | |
373 | return 1 + __sync_fetch_and_add(seq, 1); | |
374 | } | |
375 | ||
376 | static void *clock_thread_fn(void *data) | |
377 | { | |
378 | struct clock_thread *t = data; | |
379 | struct clock_entry *c; | |
380 | os_cpu_mask_t cpu_mask; | |
381 | int i; | |
382 | ||
383 | memset(&cpu_mask, 0, sizeof(cpu_mask)); | |
384 | fio_cpu_set(&cpu_mask, t->cpu); | |
385 | ||
386 | if (fio_setaffinity(gettid(), cpu_mask) == -1) { | |
387 | log_err("clock setaffinity failed\n"); | |
388 | return (void *) 1; | |
389 | } | |
390 | ||
7d11f871 | 391 | pthread_mutex_lock(&t->lock); |
b9b3498e | 392 | pthread_mutex_unlock(&t->started); |
7d11f871 JA |
393 | |
394 | c = &t->entries[0]; | |
395 | for (i = 0; i < CLOCK_ENTRIES; i++, c++) { | |
396 | uint64_t seq, tsc; | |
397 | ||
398 | c->cpu = t->cpu; | |
399 | do { | |
400 | seq = atomic64_inc_return(t->seq); | |
401 | tsc = get_cpu_clock(); | |
402 | } while (seq != *t->seq); | |
403 | ||
404 | c->seq = seq; | |
405 | c->tsc = tsc; | |
406 | } | |
407 | ||
408 | log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc); | |
409 | return NULL; | |
410 | } | |
411 | ||
412 | static int clock_cmp(const void *p1, const void *p2) | |
413 | { | |
414 | const struct clock_entry *c1 = p1; | |
415 | const struct clock_entry *c2 = p2; | |
416 | ||
b9b3498e JA |
417 | if (c1->seq == c2->seq) |
418 | log_err("cs: bug in atomic sequence!\n"); | |
419 | ||
7d11f871 JA |
420 | return c1->seq - c2->seq; |
421 | } | |
422 | ||
423 | int fio_monotonic_clocktest(void) | |
424 | { | |
425 | struct clock_thread *threads; | |
426 | unsigned int nr_cpus = cpus_online(); | |
427 | struct clock_entry *entries; | |
428 | unsigned long tentries, failed; | |
429 | uint64_t seq = 0; | |
430 | int i; | |
431 | ||
432 | threads = malloc(nr_cpus * sizeof(struct clock_thread)); | |
433 | tentries = CLOCK_ENTRIES * nr_cpus; | |
434 | entries = malloc(tentries * sizeof(struct clock_entry)); | |
435 | ||
436 | log_info("cs: Testing %u CPUs\n", nr_cpus); | |
437 | ||
438 | for (i = 0; i < nr_cpus; i++) { | |
439 | struct clock_thread *t = &threads[i]; | |
440 | ||
441 | t->cpu = i; | |
442 | t->seq = &seq; | |
443 | t->entries = &entries[i * CLOCK_ENTRIES]; | |
444 | pthread_mutex_init(&t->lock, NULL); | |
445 | pthread_mutex_init(&t->started, NULL); | |
446 | pthread_mutex_lock(&t->lock); | |
447 | pthread_create(&t->thread, NULL, clock_thread_fn, t); | |
448 | } | |
449 | ||
450 | for (i = 0; i < nr_cpus; i++) { | |
451 | struct clock_thread *t = &threads[i]; | |
452 | ||
453 | pthread_mutex_lock(&t->started); | |
454 | } | |
455 | ||
456 | for (i = 0; i < nr_cpus; i++) { | |
457 | struct clock_thread *t = &threads[i]; | |
458 | ||
459 | pthread_mutex_unlock(&t->lock); | |
460 | } | |
461 | ||
462 | for (failed = i = 0; i < nr_cpus; i++) { | |
463 | struct clock_thread *t = &threads[i]; | |
464 | void *ret; | |
465 | ||
466 | pthread_join(t->thread, &ret); | |
467 | if (ret) | |
468 | failed++; | |
469 | } | |
470 | free(threads); | |
471 | ||
472 | if (failed) { | |
473 | log_err("Clocksource test: %u threads failed\n", failed); | |
474 | goto err; | |
475 | } | |
476 | ||
477 | qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp); | |
478 | ||
479 | for (failed = i = 0; i < tentries; i++) { | |
480 | struct clock_entry *prev, *this = &entries[i]; | |
481 | ||
482 | if (!i) { | |
483 | prev = this; | |
484 | continue; | |
485 | } | |
486 | ||
487 | if (prev->tsc > this->tsc) { | |
488 | uint64_t diff = prev->tsc - this->tsc; | |
489 | ||
490 | log_info("cs: CPU clock mismatch (diff=%lu):\n", diff); | |
491 | log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq); | |
492 | log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq); | |
493 | failed++; | |
494 | } | |
495 | ||
496 | prev = this; | |
497 | } | |
498 | ||
499 | if (failed) | |
500 | log_info("cs: Failed: %lu\n", failed); | |
501 | else | |
502 | log_info("cs: Pass!\n"); | |
503 | ||
504 | err: | |
505 | free(entries); | |
506 | return !!failed; | |
507 | } | |
508 | ||
509 | #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */ | |
510 | ||
511 | int fio_monotonic_clocktest(void) | |
512 | { | |
513 | log_info("cs: current platform does not support CPU clocks\n"); | |
514 | return 0; | |
515 | } | |
516 | ||
517 | #endif |