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) && !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(const struct timeval *s, const 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(const 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(const struct timeval *s, const 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(const 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(const 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 *cthreads;
 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 #ifdef FIO_INC_DEBUG
 545         fio_debug |= 1U << FD_TIME;
 546 #endif
 547         calibrate_cpu_clock();
 548 #ifdef FIO_INC_DEBUG
 549         fio_debug &= ~(1U << FD_TIME);
 550 #endif
 551
 552         cthreads = 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 = &cthreads[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);
 567                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 568                         failed++;
 569                         nr_cpus = i;
 570                         break;
 571                 }
 572         }
 573
 574         for (i = 0; i < nr_cpus; i++) {
 575                 struct clock_thread *t = &cthreads[i];
 576
 577                 pthread_mutex_lock(&t->started);
 578         }
 579
 580         for (i = 0; i < nr_cpus; i++) {
 581                 struct clock_thread *t = &cthreads[i];
 582
 583                 pthread_mutex_unlock(&t->lock);
 584         }
 585
 586         for (i = 0; i < nr_cpus; i++) {
 587                 struct clock_thread *t = &cthreads[i];
 588                 void *ret;
 589
 590                 pthread_join(t->thread, &ret);
 591                 if (ret)
 592                         failed++;
 593         }
 594         free(cthreads);
 595
 596         if (failed) {
 597                 log_err("Clocksource test: %lu threads failed\n", failed);
 598                 goto err;
 599         }
 600
 601         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 602
 603         for (failed = i = 0; i < tentries; i++) {
 604                 this = &entries[i];
 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
 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);
 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
 629 err:
 630         free(entries);
 631         return !!failed;
 632 }
 633
 634 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 635
 636 int fio_monotonic_clocktest(void)
 637 {
 638         log_info("cs: current platform does not support CPU clocks\n");
 639         return 0;
 640 }
 641
 642 #endif