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         uint64_t last_cycles;
  24         uint64_t last_tv_valid;
  25 };
  26 #ifdef CONFIG_TLS_THREAD
  27 static __thread struct tv_valid static_tv_valid;
  28 #else
  29 static pthread_key_t tv_tls_key;
  30 #endif
  31
  32 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  33 int fio_clock_source_set = 0;
  34 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  35
  36 #ifdef FIO_DEBUG_TIME
  37
  38 #define HASH_BITS       8
  39 #define HASH_SIZE       (1 << HASH_BITS)
  40
  41 static struct flist_head hash[HASH_SIZE];
  42 static int gtod_inited;
  43
  44 struct gtod_log {
  45         struct flist_head list;
  46         void *caller;
  47         unsigned long calls;
  48 };
  49
  50 static struct gtod_log *find_hash(void *caller)
  51 {
  52         unsigned long h = hash_ptr(caller, HASH_BITS);
  53         struct flist_head *entry;
  54
  55         flist_for_each(entry, &hash[h]) {
  56                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  57                                                                         list);
  58
  59                 if (log->caller == caller)
  60                         return log;
  61         }
  62
  63         return NULL;
  64 }
  65
  66 static struct gtod_log *find_log(void *caller)
  67 {
  68         struct gtod_log *log = find_hash(caller);
  69
  70         if (!log) {
  71                 unsigned long h;
  72
  73                 log = malloc(sizeof(*log));
  74                 INIT_FLIST_HEAD(&log->list);
  75                 log->caller = caller;
  76                 log->calls = 0;
  77
  78                 h = hash_ptr(caller, HASH_BITS);
  79                 flist_add_tail(&log->list, &hash[h]);
  80         }
  81
  82         return log;
  83 }
  84
  85 static void gtod_log_caller(void *caller)
  86 {
  87         if (gtod_inited) {
  88                 struct gtod_log *log = find_log(caller);
  89
  90                 log->calls++;
  91         }
  92 }
  93
  94 static void fio_exit fio_dump_gtod(void)
  95 {
  96         unsigned long total_calls = 0;
  97         int i;
  98
  99         for (i = 0; i < HASH_SIZE; i++) {
 100                 struct flist_head *entry;
 101                 struct gtod_log *log;
 102
 103                 flist_for_each(entry, &hash[i]) {
 104                         log = flist_entry(entry, struct gtod_log, list);
 105
 106                         printf("function %p, calls %lu\n", log->caller,
 107                                                                 log->calls);
 108                         total_calls += log->calls;
 109                 }
 110         }
 111
 112         printf("Total %lu gettimeofday\n", total_calls);
 113 }
 114
 115 static void fio_init gtod_init(void)
 116 {
 117         int i;
 118
 119         for (i = 0; i < HASH_SIZE; i++)
 120                 INIT_FLIST_HEAD(&hash[i]);
 121
 122         gtod_inited = 1;
 123 }
 124
 125 #endif /* FIO_DEBUG_TIME */
 126
 127 #ifdef CONFIG_CLOCK_GETTIME
 128 static int fill_clock_gettime(struct timespec *ts)
 129 {
 130 #ifdef CONFIG_CLOCK_MONOTONIC
 131         return clock_gettime(CLOCK_MONOTONIC, ts);
 132 #else
 133         return clock_gettime(CLOCK_REALTIME, ts);
 134 #endif
 135 }
 136 #endif
 137
 138 static void __fio_gettime(struct timeval *tp)
 139 {
 140         struct tv_valid *tv;
 141
 142 #ifdef CONFIG_TLS_THREAD
 143         tv = &static_tv_valid;
 144 #else
 145         tv = pthread_getspecific(tv_tls_key);
 146 #endif
 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
 172                 t = get_cpu_clock();
 173                 if (t < tv->last_cycles && tv->last_tv_valid)
 174                         log_err("fio: CPU clock going back in time\n");
 175
 176                 tv->last_cycles = t;
 177                 tv->last_tv_valid = 1;
 178 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 179                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 180 #else
 181                 usecs = (t * inv_cycles_per_usec) / 16777216UL;
 182 #endif
 183                 tp->tv_sec = usecs / 1000000;
 184                 tp->tv_usec = usecs % 1000000;
 185                 break;
 186                 }
 187 #endif
 188         default:
 189                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 190                 break;
 191         }
 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 #ifdef FIO_DEBUG_TIME
 201         if (!caller)
 202                 caller = __builtin_return_address(0);
 203
 204         gtod_log_caller(caller);
 205 #endif
 206         if (fio_unlikely(fio_tv)) {
 207                 memcpy(tp, fio_tv, sizeof(*tp));
 208                 return;
 209         }
 210
 211         __fio_gettime(tp);
 212 }
 213
 214 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 215 static unsigned long get_cycles_per_usec(void)
 216 {
 217         struct timeval s, e;
 218         uint64_t c_s, c_e;
 219         enum fio_cs old_cs = fio_clock_source;
 220
 221 #ifdef CONFIG_CLOCK_GETTIME
 222         fio_clock_source = CS_CGETTIME;
 223 #else
 224         fio_clock_source = CS_GTOD;
 225 #endif
 226         __fio_gettime(&s);
 227
 228         c_s = get_cpu_clock();
 229         do {
 230                 uint64_t elapsed;
 231
 232                 __fio_gettime(&e);
 233
 234                 elapsed = utime_since(&s, &e);
 235                 if (elapsed >= 1280) {
 236                         c_e = get_cpu_clock();
 237                         break;
 238                 }
 239         } while (1);
 240
 241         fio_clock_source = old_cs;
 242         return (c_e - c_s + 127) >> 7;
 243 }
 244
 245 #define NR_TIME_ITERS   50
 246
 247 static int calibrate_cpu_clock(void)
 248 {
 249         double delta, mean, S;
 250         uint64_t avg, cycles[NR_TIME_ITERS];
 251         int i, samples;
 252
 253         cycles[0] = get_cycles_per_usec();
 254         S = delta = mean = 0.0;
 255         for (i = 0; i < NR_TIME_ITERS; i++) {
 256                 cycles[i] = get_cycles_per_usec();
 257                 delta = cycles[i] - mean;
 258                 if (delta) {
 259                         mean += delta / (i + 1.0);
 260                         S += delta * (cycles[i] - mean);
 261                 }
 262         }
 263
 264         /*
 265          * The most common platform clock breakage is returning zero
 266          * indefinitely. Check for that and return failure.
 267          */
 268         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 269                 return 1;
 270
 271         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 272
 273         samples = avg = 0;
 274         for (i = 0; i < NR_TIME_ITERS; i++) {
 275                 double this = cycles[i];
 276
 277                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 278                         continue;
 279                 samples++;
 280                 avg += this;
 281         }
 282
 283         S /= (double) NR_TIME_ITERS;
 284         mean /= 10.0;
 285
 286         for (i = 0; i < NR_TIME_ITERS; i++)
 287                 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
 288                                         (unsigned long long) cycles[i] / 10);
 289
 290         avg /= samples;
 291         avg = (avg + 5) / 10;
 292         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 293         dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
 294
 295         cycles_per_usec = avg;
 296         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
 297         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
 298         return 0;
 299 }
 300 #else
 301 static int calibrate_cpu_clock(void)
 302 {
 303 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 304         return 0;
 305 #else
 306         return 1;
 307 #endif
 308 }
 309 #endif // ARCH_HAVE_CPU_CLOCK
 310
 311 #ifndef CONFIG_TLS_THREAD
 312 void fio_local_clock_init(int is_thread)
 313 {
 314         struct tv_valid *t;
 315
 316         t = calloc(1, sizeof(*t));
 317         if (pthread_setspecific(tv_tls_key, t)) {
 318                 log_err("fio: can't set TLS key\n");
 319                 assert(0);
 320         }
 321 }
 322
 323 static void kill_tv_tls_key(void *data)
 324 {
 325         free(data);
 326 }
 327 #else
 328 void fio_local_clock_init(int is_thread)
 329 {
 330 }
 331 #endif
 332
 333 void fio_clock_init(void)
 334 {
 335         if (fio_clock_source == fio_clock_source_inited)
 336                 return;
 337
 338 #ifndef CONFIG_TLS_THREAD
 339         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 340                 log_err("fio: can't create TLS key\n");
 341 #endif
 342
 343         fio_clock_source_inited = fio_clock_source;
 344
 345         if (calibrate_cpu_clock())
 346                 tsc_reliable = 0;
 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(const struct timeval *s, const 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(const 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(const struct timeval *s, const 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(const 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(const 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         uint32_t seq;
 433         uint32_t cpu;
 434         uint64_t tsc;
 435 };
 436
 437 struct clock_thread {
 438         pthread_t thread;
 439         int cpu;
 440         pthread_mutex_t lock;
 441         pthread_mutex_t started;
 442         uint32_t *seq;
 443         struct clock_entry *entries;
 444 };
 445
 446 static inline uint32_t atomic32_inc_return(uint32_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         uint32_t last_seq;
 457         int i;
 458
 459         memset(&cpu_mask, 0, sizeof(cpu_mask));
 460         fio_cpu_set(&cpu_mask, t->cpu);
 461
 462         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 463                 log_err("clock setaffinity failed\n");
 464                 return (void *) 1;
 465         }
 466
 467         pthread_mutex_lock(&t->lock);
 468         pthread_mutex_unlock(&t->started);
 469
 470         last_seq = 0;
 471         c = &t->entries[0];
 472         for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
 473                 uint32_t seq;
 474                 uint64_t tsc;
 475
 476                 c->cpu = t->cpu;
 477                 do {
 478                         seq = atomic32_inc_return(t->seq);
 479                         if (seq < last_seq)
 480                                 break;
 481                         tsc = get_cpu_clock();
 482                 } while (seq != *t->seq);
 483
 484                 c->seq = seq;
 485                 c->tsc = tsc;
 486         }
 487
 488         log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
 489                 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
 490
 491         /*
 492          * The most common platform clock breakage is returning zero
 493          * indefinitely. Check for that and return failure.
 494          */
 495         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 496                 return (void *) 1;
 497
 498         return NULL;
 499 }
 500
 501 static int clock_cmp(const void *p1, const void *p2)
 502 {
 503         const struct clock_entry *c1 = p1;
 504         const struct clock_entry *c2 = p2;
 505
 506         if (c1->seq == c2->seq)
 507                 log_err("cs: bug in atomic sequence!\n");
 508
 509         return c1->seq - c2->seq;
 510 }
 511
 512 int fio_monotonic_clocktest(void)
 513 {
 514         struct clock_thread *cthreads;
 515         unsigned int nr_cpus = cpus_online();
 516         struct clock_entry *entries;
 517         unsigned long tentries, failed = 0;
 518         struct clock_entry *prev, *this;
 519         uint32_t seq = 0;
 520         unsigned int i;
 521
 522         log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 523
 524 #ifdef FIO_INC_DEBUG
 525         fio_debug |= 1U << FD_TIME;
 526 #endif
 527         calibrate_cpu_clock();
 528 #ifdef FIO_INC_DEBUG
 529         fio_debug &= ~(1U << FD_TIME);
 530 #endif
 531
 532         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 533         tentries = CLOCK_ENTRIES * nr_cpus;
 534         entries = malloc(tentries * sizeof(struct clock_entry));
 535
 536         log_info("cs: Testing %u CPUs\n", nr_cpus);
 537
 538         for (i = 0; i < nr_cpus; i++) {
 539                 struct clock_thread *t = &cthreads[i];
 540
 541                 t->cpu = i;
 542                 t->seq = &seq;
 543                 t->entries = &entries[i * CLOCK_ENTRIES];
 544                 pthread_mutex_init(&t->lock, NULL);
 545                 pthread_mutex_init(&t->started, NULL);
 546                 pthread_mutex_lock(&t->lock);
 547                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 548                         failed++;
 549                         nr_cpus = i;
 550                         break;
 551                 }
 552         }
 553
 554         for (i = 0; i < nr_cpus; i++) {
 555                 struct clock_thread *t = &cthreads[i];
 556
 557                 pthread_mutex_lock(&t->started);
 558         }
 559
 560         for (i = 0; i < nr_cpus; i++) {
 561                 struct clock_thread *t = &cthreads[i];
 562
 563                 pthread_mutex_unlock(&t->lock);
 564         }
 565
 566         for (i = 0; i < nr_cpus; i++) {
 567                 struct clock_thread *t = &cthreads[i];
 568                 void *ret;
 569
 570                 pthread_join(t->thread, &ret);
 571                 if (ret)
 572                         failed++;
 573         }
 574         free(cthreads);
 575
 576         if (failed) {
 577                 log_err("Clocksource test: %lu threads failed\n", failed);
 578                 goto err;
 579         }
 580
 581         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 582
 583         for (failed = i = 0; i < tentries; i++) {
 584                 this = &entries[i];
 585
 586                 if (!i) {
 587                         prev = this;
 588                         continue;
 589                 }
 590
 591                 if (prev->tsc > this->tsc) {
 592                         uint64_t diff = prev->tsc - this->tsc;
 593
 594                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 595                                                 (unsigned long long) diff);
 596                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 597                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 598                         failed++;
 599                 }
 600
 601                 prev = this;
 602         }
 603
 604         if (failed)
 605                 log_info("cs: Failed: %lu\n", failed);
 606         else
 607                 log_info("cs: Pass!\n");
 608
 609 err:
 610         free(entries);
 611         return !!failed;
 612 }
 613
 614 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 615
 616 int fio_monotonic_clocktest(void)
 617 {
 618         log_info("cs: current platform does not support CPU clocks\n");
 619         return 0;
 620 }
 621
 622 #endif