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) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC) 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 __thread struct tv_valid 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 static 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 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC 181 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC; 182 #else 183 usecs = (t * inv_cycles_per_usec) / 16777216UL; 184 #endif 185 tp->tv_sec = usecs / 1000000; 186 tp->tv_usec = usecs % 1000000; 187 break; 188 } 189 #endif 190 default: 191 log_err("fio: invalid clock source %d\n", fio_clock_source); 192 break; 193 } 194 195 return tv; 196 } 197 198 #ifdef FIO_DEBUG_TIME 199 void fio_gettime(struct timeval *tp, void *caller) 200 #else 201 void 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 212 if (fio_unlikely(fio_tv)) { 213 memcpy(tp, fio_tv, sizeof(*tp)); 214 return; 215 } 216 217 tv = __fio_gettime(tp); 218 219 /* 220 * If Linux is using the tsc clock on non-synced processors, 221 * sometimes time can appear to drift backwards. Fix that up. 222 */ 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; 230 } 231 tv->last_tv_valid = 1; 232 memcpy(&tv->last_tv, tp, sizeof(*tp)); 233 } 234 } 235 236 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC) 237 static unsigned long get_cycles_per_usec(void) 238 { 239 struct timeval s, e; 240 uint64_t c_s, c_e; 241 enum fio_cs old_cs = fio_clock_source; 242 243 #ifdef CONFIG_CLOCK_GETTIME 244 fio_clock_source = CS_CGETTIME; 245 #else 246 fio_clock_source = CS_GTOD; 247 #endif 248 __fio_gettime(&s); 249 250 c_s = get_cpu_clock(); 251 do { 252 uint64_t elapsed; 253 254 __fio_gettime(&e); 255 256 elapsed = utime_since(&s, &e); 257 if (elapsed >= 1280) { 258 c_e = get_cpu_clock(); 259 break; 260 } 261 } while (1); 262 263 fio_clock_source = old_cs; 264 return (c_e - c_s + 127) >> 7; 265 } 266 267 #define NR_TIME_ITERS 50 268 269 static int calibrate_cpu_clock(void) 270 { 271 double delta, mean, S; 272 uint64_t avg, cycles[NR_TIME_ITERS]; 273 int i, samples; 274 275 cycles[0] = get_cycles_per_usec(); 276 S = delta = mean = 0.0; 277 for (i = 0; i < NR_TIME_ITERS; i++) { 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 } 285 286 /* 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; 292 293 S = sqrt(S / (NR_TIME_ITERS - 1.0)); 294 295 samples = avg = 0; 296 for (i = 0; i < NR_TIME_ITERS; i++) { 297 double this = cycles[i]; 298 299 if ((fmax(this, mean) - fmin(this, mean)) > S) 300 continue; 301 samples++; 302 avg += this; 303 } 304 305 S /= (double) NR_TIME_ITERS; 306 mean /= 10.0; 307 308 for (i = 0; i < NR_TIME_ITERS; i++) 309 dprint(FD_TIME, "cycles[%d]=%llu\n", i, 310 (unsigned long long) cycles[i] / 10); 311 312 avg /= samples; 313 avg = (avg + 5) / 10; 314 dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg); 315 dprint(FD_TIME, "mean=%f, S=%f\n", mean, S); 316 317 cycles_per_usec = avg; 318 inv_cycles_per_usec = 16777216UL / cycles_per_usec; 319 dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec); 320 return 0; 321 } 322 #else 323 static int calibrate_cpu_clock(void) 324 { 325 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC 326 return 0; 327 #else 328 return 1; 329 #endif 330 } 331 #endif // ARCH_HAVE_CPU_CLOCK 332 333 #ifndef CONFIG_TLS_THREAD 334 void fio_local_clock_init(int is_thread) 335 { 336 struct tv_valid *t; 337 338 t = calloc(1, sizeof(*t)); 339 if (pthread_setspecific(tv_tls_key, t)) 340 log_err("fio: can't set TLS key\n"); 341 } 342 343 static void kill_tv_tls_key(void *data) 344 { 345 free(data); 346 } 347 #else 348 void fio_local_clock_init(int is_thread) 349 { 350 } 351 #endif 352 353 void fio_clock_init(void) 354 { 355 if (fio_clock_source == fio_clock_source_inited) 356 return; 357 358 #ifndef CONFIG_TLS_THREAD 359 if (pthread_key_create(&tv_tls_key, kill_tv_tls_key)) 360 log_err("fio: can't create TLS key\n"); 361 #endif 362 363 fio_clock_source_inited = fio_clock_source; 364 365 if (calibrate_cpu_clock()) 366 tsc_reliable = 0; 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"); 378 } 379 380 uint64_t utime_since(struct timeval *s, struct timeval *e) 381 { 382 long sec, usec; 383 uint64_t ret; 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; 401 } 402 403 uint64_t utime_since_now(struct timeval *s) 404 { 405 struct timeval t; 406 407 fio_gettime(&t, NULL); 408 return utime_since(s, &t); 409 } 410 411 uint64_t mtime_since(struct timeval *s, struct timeval *e) 412 { 413 long sec, usec, ret; 414 415 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; 420 } 421 422 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; 430 } 431 432 uint64_t mtime_since_now(struct timeval *s) 433 { 434 struct timeval t; 435 void *p = __builtin_return_address(0); 436 437 fio_gettime(&t, p); 438 return mtime_since(s, &t); 439 } 440 441 uint64_t time_since_now(struct timeval *s) 442 { 443 return mtime_since_now(s) / 1000; 444 } 445 446 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \ 447 defined(CONFIG_SFAA) 448 449 #define CLOCK_ENTRIES 100000 450 451 struct clock_entry { 452 uint32_t seq; 453 uint32_t cpu; 454 uint64_t tsc; 455 }; 456 457 struct clock_thread { 458 pthread_t thread; 459 int cpu; 460 pthread_mutex_t lock; 461 pthread_mutex_t started; 462 uint32_t *seq; 463 struct clock_entry *entries; 464 }; 465 466 static inline uint32_t atomic32_inc_return(uint32_t *seq) 467 { 468 return 1 + __sync_fetch_and_add(seq, 1); 469 } 470 471 static 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; 476 uint32_t last_seq; 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 } 486 487 pthread_mutex_lock(&t->lock); 488 pthread_mutex_unlock(&t->started); 489 490 last_seq = 0; 491 c = &t->entries[0]; 492 for (i = 0; i < CLOCK_ENTRIES; i++, c++) { 493 uint32_t seq; 494 uint64_t tsc; 495 496 c->cpu = t->cpu; 497 do { 498 seq = atomic32_inc_return(t->seq); 499 if (seq < last_seq) 500 break; 501 tsc = get_cpu_clock(); 502 } while (seq != *t->seq); 503 504 c->seq = seq; 505 c->tsc = tsc; 506 } 507 508 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu, 509 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc); 510 511 /* 512 * The most common platform clock breakage is returning zero 513 * indefinitely. Check for that and return failure. 514 */ 515 if (!t->entries[i - 1].tsc && !t->entries[0].tsc) 516 return (void *) 1; 517 518 return NULL; 519 } 520 521 static int clock_cmp(const void *p1, const void *p2) 522 { 523 const struct clock_entry *c1 = p1; 524 const struct clock_entry *c2 = p2; 525 526 if (c1->seq == c2->seq) 527 log_err("cs: bug in atomic sequence!\n"); 528 529 return c1->seq - c2->seq; 530 } 531 532 int fio_monotonic_clocktest(void) 533 { 534 struct clock_thread *threads; 535 unsigned int nr_cpus = cpus_online(); 536 struct clock_entry *entries; 537 unsigned long tentries, failed = 0; 538 struct clock_entry *prev, *this; 539 uint32_t seq = 0; 540 unsigned int i; 541 542 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no"); 543 544 fio_debug |= 1U << FD_TIME; 545 calibrate_cpu_clock(); 546 fio_debug &= ~(1U << FD_TIME); 547 548 threads = malloc(nr_cpus * sizeof(struct clock_thread)); 549 tentries = CLOCK_ENTRIES * nr_cpus; 550 entries = malloc(tentries * sizeof(struct clock_entry)); 551 552 log_info("cs: Testing %u CPUs\n", nr_cpus); 553 554 for (i = 0; i < nr_cpus; i++) { 555 struct clock_thread *t = &threads[i]; 556 557 t->cpu = i; 558 t->seq = &seq; 559 t->entries = &entries[i * CLOCK_ENTRIES]; 560 pthread_mutex_init(&t->lock, NULL); 561 pthread_mutex_init(&t->started, NULL); 562 pthread_mutex_lock(&t->lock); 563 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) { 564 failed++; 565 nr_cpus = i; 566 break; 567 } 568 } 569 570 for (i = 0; i < nr_cpus; i++) { 571 struct clock_thread *t = &threads[i]; 572 573 pthread_mutex_lock(&t->started); 574 } 575 576 for (i = 0; i < nr_cpus; i++) { 577 struct clock_thread *t = &threads[i]; 578 579 pthread_mutex_unlock(&t->lock); 580 } 581 582 for (i = 0; i < nr_cpus; i++) { 583 struct clock_thread *t = &threads[i]; 584 void *ret; 585 586 pthread_join(t->thread, &ret); 587 if (ret) 588 failed++; 589 } 590 free(threads); 591 592 if (failed) { 593 log_err("Clocksource test: %lu threads failed\n", failed); 594 goto err; 595 } 596 597 qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp); 598 599 for (failed = i = 0; i < tentries; i++) { 600 this = &entries[i]; 601 602 if (!i) { 603 prev = this; 604 continue; 605 } 606 607 if (prev->tsc > this->tsc) { 608 uint64_t diff = prev->tsc - this->tsc; 609 610 log_info("cs: CPU clock mismatch (diff=%llu):\n", 611 (unsigned long long) diff); 612 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq); 613 log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq); 614 failed++; 615 } 616 617 prev = this; 618 } 619 620 if (failed) 621 log_info("cs: Failed: %lu\n", failed); 622 else 623 log_info("cs: Pass!\n"); 624 625 err: 626 free(entries); 627 return !!failed; 628 } 629 630 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */ 631 632 int fio_monotonic_clocktest(void) 633 { 634 log_info("cs: current platform does not support CPU clocks\n"); 635 return 0; 636 } 637 638 #endif 639
