gettime.c

   1 /*
   2  * Clock functions
   3  */
   4
   5 #include <math.h>
   6
   7 #include "fio.h"
   8 #include "os/os.h"
   9
  10 #if defined(ARCH_HAVE_CPU_CLOCK)
  11 #ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
  12 static unsigned long long cycles_per_msec;
  13 static unsigned long long cycles_start;
  14 static unsigned long long clock_mult;
  15 static unsigned long long max_cycles_mask;
  16 static unsigned long long nsecs_for_max_cycles;
  17 static unsigned int clock_shift;
  18 static unsigned int max_cycles_shift;
  19 #define MAX_CLOCK_SEC 60*60
  20 #endif
  21 #ifdef ARCH_CPU_CLOCK_WRAPS
  22 static unsigned int cycles_wrap;
  23 #endif
  24 #endif
  25 bool tsc_reliable = false;
  26
  27 struct tv_valid {
  28         int warned;
  29 };
  30 #ifdef ARCH_HAVE_CPU_CLOCK
  31 #ifdef CONFIG_TLS_THREAD
  32 static __thread struct tv_valid static_tv_valid;
  33 #else
  34 static pthread_key_t tv_tls_key;
  35 #endif
  36 #endif
  37
  38 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  39 int fio_clock_source_set = 0;
  40 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  41
  42 #ifdef FIO_DEBUG_TIME
  43
  44 #define HASH_BITS       8
  45 #define HASH_SIZE       (1 << HASH_BITS)
  46
  47 static struct flist_head hash[HASH_SIZE];
  48 static int gtod_inited;
  49
  50 struct gtod_log {
  51         struct flist_head list;
  52         void *caller;
  53         unsigned long calls;
  54 };
  55
  56 static struct gtod_log *find_hash(void *caller)
  57 {
  58         unsigned long h = hash_ptr(caller, HASH_BITS);
  59         struct flist_head *entry;
  60
  61         flist_for_each(entry, &hash[h]) {
  62                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  63                                                                         list);
  64
  65                 if (log->caller == caller)
  66                         return log;
  67         }
  68
  69         return NULL;
  70 }
  71
  72 static void inc_caller(void *caller)
  73 {
  74         struct gtod_log *log = find_hash(caller);
  75
  76         if (!log) {
  77                 unsigned long h;
  78
  79                 log = malloc(sizeof(*log));
  80                 INIT_FLIST_HEAD(&log->list);
  81                 log->caller = caller;
  82                 log->calls = 0;
  83
  84                 h = hash_ptr(caller, HASH_BITS);
  85                 flist_add_tail(&log->list, &hash[h]);
  86         }
  87
  88         log->calls++;
  89 }
  90
  91 static void gtod_log_caller(void *caller)
  92 {
  93         if (gtod_inited)
  94                 inc_caller(caller);
  95 }
  96
  97 static void fio_exit fio_dump_gtod(void)
  98 {
  99         unsigned long total_calls = 0;
 100         int i;
 101
 102         for (i = 0; i < HASH_SIZE; i++) {
 103                 struct flist_head *entry;
 104                 struct gtod_log *log;
 105
 106                 flist_for_each(entry, &hash[i]) {
 107                         log = flist_entry(entry, struct gtod_log, list);
 108
 109                         printf("function %p, calls %lu\n", log->caller,
 110                                                                 log->calls);
 111                         total_calls += log->calls;
 112                 }
 113         }
 114
 115         printf("Total %lu gettimeofday\n", total_calls);
 116 }
 117
 118 static void fio_init gtod_init(void)
 119 {
 120         int i;
 121
 122         for (i = 0; i < HASH_SIZE; i++)
 123                 INIT_FLIST_HEAD(&hash[i]);
 124
 125         gtod_inited = 1;
 126 }
 127
 128 #endif /* FIO_DEBUG_TIME */
 129
 130 #ifdef CONFIG_CLOCK_GETTIME
 131 static int fill_clock_gettime(struct timespec *ts)
 132 {
 133 #if defined(CONFIG_CLOCK_MONOTONIC_RAW)
 134         return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
 135 #elif defined(CONFIG_CLOCK_MONOTONIC)
 136         return clock_gettime(CLOCK_MONOTONIC, ts);
 137 #else
 138         return clock_gettime(CLOCK_REALTIME, ts);
 139 #endif
 140 }
 141 #endif
 142
 143 static void __fio_gettime(struct timespec *tp)
 144 {
 145         switch (fio_clock_source) {
 146 #ifdef CONFIG_GETTIMEOFDAY
 147         case CS_GTOD: {
 148                 struct timeval tv;
 149                 gettimeofday(&tv, NULL);
 150
 151                 tp->tv_sec = tv.tv_sec;
 152                 tp->tv_nsec = tv.tv_usec * 1000;
 153                 break;
 154                 }
 155 #endif
 156 #ifdef CONFIG_CLOCK_GETTIME
 157         case CS_CGETTIME: {
 158                 if (fill_clock_gettime(tp) < 0) {
 159                         log_err("fio: clock_gettime fails\n");
 160                         assert(0);
 161                 }
 162                 break;
 163                 }
 164 #endif
 165 #ifdef ARCH_HAVE_CPU_CLOCK
 166         case CS_CPUCLOCK: {
 167                 uint64_t nsecs, t, multiples;
 168                 struct tv_valid *tv;
 169
 170 #ifdef CONFIG_TLS_THREAD
 171                 tv = &static_tv_valid;
 172 #else
 173                 tv = pthread_getspecific(tv_tls_key);
 174 #endif
 175
 176                 t = get_cpu_clock();
 177 #ifdef ARCH_CPU_CLOCK_WRAPS
 178                 if (t < cycles_start && !cycles_wrap)
 179                         cycles_wrap = 1;
 180                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
 181                         log_err("fio: double CPU clock wrap\n");
 182                         tv->warned = 1;
 183                 }
 184 #endif
 185 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 186                 nsecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC * 1000;
 187 #else
 188                 t -= cycles_start;
 189                 multiples = t >> max_cycles_shift;
 190                 nsecs = multiples * nsecs_for_max_cycles;
 191                 nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift;
 192 #endif
 193                 tp->tv_sec = nsecs / 1000000000ULL;
 194                 tp->tv_nsec = nsecs % 1000000000ULL;
 195                 break;
 196                 }
 197 #endif
 198         default:
 199                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 200                 break;
 201         }
 202 }
 203
 204 #ifdef FIO_DEBUG_TIME
 205 void fio_gettime(struct timespec *tp, void *caller)
 206 #else
 207 void fio_gettime(struct timespec *tp, void fio_unused *caller)
 208 #endif
 209 {
 210 #ifdef FIO_DEBUG_TIME
 211         if (!caller)
 212                 caller = __builtin_return_address(0);
 213
 214         gtod_log_caller(caller);
 215 #endif
 216         if (fio_unlikely(fio_gettime_offload(tp)))
 217                 return;
 218
 219         __fio_gettime(tp);
 220 }
 221
 222 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 223 static unsigned long get_cycles_per_msec(void)
 224 {
 225         struct timespec s, e;
 226         uint64_t c_s, c_e;
 227         enum fio_cs old_cs = fio_clock_source;
 228         uint64_t elapsed;
 229
 230 #ifdef CONFIG_CLOCK_GETTIME
 231         fio_clock_source = CS_CGETTIME;
 232 #else
 233         fio_clock_source = CS_GTOD;
 234 #endif
 235         __fio_gettime(&s);
 236
 237         c_s = get_cpu_clock();
 238         do {
 239                 __fio_gettime(&e);
 240                 c_e = get_cpu_clock();
 241
 242                 elapsed = utime_since(&s, &e);
 243                 if (elapsed >= 1280)
 244                         break;
 245         } while (1);
 246
 247         fio_clock_source = old_cs;
 248         return (c_e - c_s) * 1000 / elapsed;
 249 }
 250
 251 #define NR_TIME_ITERS   50
 252
 253 static int calibrate_cpu_clock(void)
 254 {
 255         double delta, mean, S;
 256         uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
 257         int i, samples, sft = 0;
 258         unsigned long long tmp, max_ticks, max_mult;
 259
 260         cycles[0] = get_cycles_per_msec();
 261         S = delta = mean = 0.0;
 262         for (i = 0; i < NR_TIME_ITERS; i++) {
 263                 cycles[i] = get_cycles_per_msec();
 264                 delta = cycles[i] - mean;
 265                 if (delta) {
 266                         mean += delta / (i + 1.0);
 267                         S += delta * (cycles[i] - mean);
 268                 }
 269         }
 270
 271         /*
 272          * The most common platform clock breakage is returning zero
 273          * indefinitely. Check for that and return failure.
 274          */
 275         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 276                 return 1;
 277
 278         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 279
 280         minc = -1ULL;
 281         maxc = samples = avg = 0;
 282         for (i = 0; i < NR_TIME_ITERS; i++) {
 283                 double this = cycles[i];
 284
 285                 minc = min(cycles[i], minc);
 286                 maxc = max(cycles[i], maxc);
 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
 296         for (i = 0; i < NR_TIME_ITERS; i++)
 297                 dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
 298
 299         avg /= samples;
 300         cycles_per_msec = avg;
 301         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 302         dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
 303                         (unsigned long long) minc,
 304                         (unsigned long long) maxc, mean, S);
 305
 306         max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL;
 307         max_mult = ULLONG_MAX / max_ticks;
 308         dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, "
 309                         "max_mult=%llu\n", max_ticks,
 310                         __builtin_clzll(max_ticks), max_mult);
 311
 312         /*
 313          * Find the largest shift count that will produce
 314          * a multiplier that does not exceed max_mult
 315          */
 316         tmp = max_mult * cycles_per_msec / 1000000;
 317         while (tmp > 1) {
 318                 tmp >>= 1;
 319                 sft++;
 320                 dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft);
 321         }
 322
 323         clock_shift = sft;
 324         clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec;
 325         dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift,
 326                                                         clock_mult);
 327
 328         /*
 329          * Find the greatest power of 2 clock ticks that is less than the
 330          * ticks in MAX_CLOCK_SEC_2STAGE
 331          */
 332         max_cycles_shift = max_cycles_mask = 0;
 333         tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec;
 334         dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp,
 335                                                         max_cycles_shift);
 336         while (tmp > 1) {
 337                 tmp >>= 1;
 338                 max_cycles_shift++;
 339                 dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
 340         }
 341         /*
 342          * if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec
 343          * here we will have a discontinuity every
 344          * (1ULL << max_cycles_shift) cycles
 345          */
 346         nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult)
 347                                         >> clock_shift;
 348
 349         /* Use a bitmask to calculate ticks % (1ULL << max_cycles_shift) */
 350         for (tmp = 0; tmp < max_cycles_shift; tmp++)
 351                 max_cycles_mask |= 1ULL << tmp;
 352
 353         dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, "
 354                         "nsecs_for_max_cycles=%llu, "
 355                         "max_cycles_mask=%016llx\n",
 356                         max_cycles_shift, (1ULL << max_cycles_shift),
 357                         nsecs_for_max_cycles, max_cycles_mask);
 358
 359         cycles_start = get_cpu_clock();
 360         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
 361         return 0;
 362 }
 363 #else
 364 static int calibrate_cpu_clock(void)
 365 {
 366 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 367         return 0;
 368 #else
 369         return 1;
 370 #endif
 371 }
 372 #endif // ARCH_HAVE_CPU_CLOCK
 373
 374 #ifndef CONFIG_TLS_THREAD
 375 void fio_local_clock_init(void)
 376 {
 377         struct tv_valid *t;
 378
 379         t = calloc(1, sizeof(*t));
 380         if (pthread_setspecific(tv_tls_key, t)) {
 381                 log_err("fio: can't set TLS key\n");
 382                 assert(0);
 383         }
 384 }
 385
 386 static void kill_tv_tls_key(void *data)
 387 {
 388         free(data);
 389 }
 390 #else
 391 void fio_local_clock_init(void)
 392 {
 393 }
 394 #endif
 395
 396 void fio_clock_init(void)
 397 {
 398         if (fio_clock_source == fio_clock_source_inited)
 399                 return;
 400
 401 #ifndef CONFIG_TLS_THREAD
 402         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 403                 log_err("fio: can't create TLS key\n");
 404 #endif
 405
 406         fio_clock_source_inited = fio_clock_source;
 407
 408         if (calibrate_cpu_clock())
 409                 tsc_reliable = false;
 410
 411         /*
 412          * If the arch sets tsc_reliable != 0, then it must be good enough
 413          * to use as THE clock source. For x86 CPUs, this means the TSC
 414          * runs at a constant rate and is synced across CPU cores.
 415          */
 416         if (tsc_reliable) {
 417                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
 418                         fio_clock_source = CS_CPUCLOCK;
 419         } else if (fio_clock_source == CS_CPUCLOCK)
 420                 log_info("fio: clocksource=cpu may not be reliable\n");
 421         dprint(FD_TIME, "gettime: clocksource=%d\n", (int) fio_clock_source);
 422 }
 423
 424 uint64_t ntime_since(const struct timespec *s, const struct timespec *e)
 425 {
 426        int64_t sec, nsec;
 427
 428        sec = e->tv_sec - s->tv_sec;
 429        nsec = e->tv_nsec - s->tv_nsec;
 430        if (sec > 0 && nsec < 0) {
 431                sec--;
 432                nsec += 1000000000LL;
 433        }
 434
 435        /*
 436         * time warp bug on some kernels?
 437         */
 438        if (sec < 0 || (sec == 0 && nsec < 0))
 439                return 0;
 440
 441        return nsec + (sec * 1000000000LL);
 442 }
 443
 444 uint64_t ntime_since_now(const struct timespec *s)
 445 {
 446         struct timespec now;
 447
 448         fio_gettime(&now, NULL);
 449         return ntime_since(s, &now);
 450 }
 451
 452 uint64_t utime_since(const struct timespec *s, const struct timespec *e)
 453 {
 454         int64_t sec, usec;
 455
 456         sec = e->tv_sec - s->tv_sec;
 457         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 458         if (sec > 0 && usec < 0) {
 459                 sec--;
 460                 usec += 1000000;
 461         }
 462
 463         /*
 464          * time warp bug on some kernels?
 465          */
 466         if (sec < 0 || (sec == 0 && usec < 0))
 467                 return 0;
 468
 469         return usec + (sec * 1000000);
 470 }
 471
 472 uint64_t utime_since_now(const struct timespec *s)
 473 {
 474         struct timespec t;
 475 #ifdef FIO_DEBUG_TIME
 476         void *p = __builtin_return_address(0);
 477
 478         fio_gettime(&t, p);
 479 #else
 480         fio_gettime(&t, NULL);
 481 #endif
 482
 483         return utime_since(s, &t);
 484 }
 485
 486 uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
 487 {
 488         int64_t sec, usec;
 489
 490         sec = e->tv_sec - s->tv_sec;
 491         usec = (e->tv_usec - s->tv_usec);
 492         if (sec > 0 && usec < 0) {
 493                 sec--;
 494                 usec += 1000000;
 495         }
 496
 497         if (sec < 0 || (sec == 0 && usec < 0))
 498                 return 0;
 499
 500         sec *= 1000;
 501         usec /= 1000;
 502         return sec + usec;
 503 }
 504
 505 uint64_t mtime_since_now(const struct timespec *s)
 506 {
 507         struct timespec t;
 508 #ifdef FIO_DEBUG_TIME
 509         void *p = __builtin_return_address(0);
 510
 511         fio_gettime(&t, p);
 512 #else
 513         fio_gettime(&t, NULL);
 514 #endif
 515
 516         return mtime_since(s, &t);
 517 }
 518
 519 uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
 520 {
 521         int64_t sec, usec;
 522
 523         sec = e->tv_sec - s->tv_sec;
 524         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 525         if (sec > 0 && usec < 0) {
 526                 sec--;
 527                 usec += 1000000;
 528         }
 529
 530         if (sec < 0 || (sec == 0 && usec < 0))
 531                 return 0;
 532
 533         sec *= 1000;
 534         usec /= 1000;
 535         return sec + usec;
 536 }
 537
 538 uint64_t time_since_now(const struct timespec *s)
 539 {
 540         return mtime_since_now(s) / 1000;
 541 }
 542
 543 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 544     defined(CONFIG_SYNC_SYNC) && defined(CONFIG_CMP_SWAP)
 545
 546 #define CLOCK_ENTRIES_DEBUG     100000
 547 #define CLOCK_ENTRIES_TEST      1000
 548
 549 struct clock_entry {
 550         uint32_t seq;
 551         uint32_t cpu;
 552         uint64_t tsc;
 553 };
 554
 555 struct clock_thread {
 556         pthread_t thread;
 557         int cpu;
 558         int debug;
 559         struct fio_sem lock;
 560         unsigned long nr_entries;
 561         uint32_t *seq;
 562         struct clock_entry *entries;
 563 };
 564
 565 static inline uint32_t atomic32_compare_and_swap(uint32_t *ptr, uint32_t old,
 566                                                  uint32_t new)
 567 {
 568         return __sync_val_compare_and_swap(ptr, old, new);
 569 }
 570
 571 static void *clock_thread_fn(void *data)
 572 {
 573         struct clock_thread *t = data;
 574         struct clock_entry *c;
 575         os_cpu_mask_t cpu_mask;
 576         unsigned long long first;
 577         int i;
 578
 579         if (fio_cpuset_init(&cpu_mask)) {
 580                 int __err = errno;
 581
 582                 log_err("clock cpuset init failed: %s\n", strerror(__err));
 583                 goto err_out;
 584         }
 585
 586         fio_cpu_set(&cpu_mask, t->cpu);
 587
 588         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 589                 int __err = errno;
 590
 591                 log_err("clock setaffinity failed: %s\n", strerror(__err));
 592                 goto err;
 593         }
 594
 595         fio_sem_down(&t->lock);
 596
 597         first = get_cpu_clock();
 598         c = &t->entries[0];
 599         for (i = 0; i < t->nr_entries; i++, c++) {
 600                 uint32_t seq;
 601                 uint64_t tsc;
 602
 603                 c->cpu = t->cpu;
 604                 do {
 605                         seq = *t->seq;
 606                         if (seq == UINT_MAX)
 607                                 break;
 608                         __sync_synchronize();
 609                         tsc = get_cpu_clock();
 610                 } while (seq != atomic32_compare_and_swap(t->seq, seq, seq + 1));
 611
 612                 if (seq == UINT_MAX)
 613                         break;
 614
 615                 c->seq = seq;
 616                 c->tsc = tsc;
 617         }
 618
 619         if (t->debug) {
 620                 unsigned long long clocks;
 621
 622                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
 623                 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
 624                                                         clocks, first);
 625         }
 626
 627         /*
 628          * The most common platform clock breakage is returning zero
 629          * indefinitely. Check for that and return failure.
 630          */
 631         if (i > 1 && !t->entries[i - 1].tsc && !t->entries[0].tsc)
 632                 goto err;
 633
 634         fio_cpuset_exit(&cpu_mask);
 635         return NULL;
 636 err:
 637         fio_cpuset_exit(&cpu_mask);
 638 err_out:
 639         return (void *) 1;
 640 }
 641
 642 static int clock_cmp(const void *p1, const void *p2)
 643 {
 644         const struct clock_entry *c1 = p1;
 645         const struct clock_entry *c2 = p2;
 646
 647         if (c1->seq == c2->seq)
 648                 log_err("cs: bug in atomic sequence!\n");
 649
 650         return c1->seq - c2->seq;
 651 }
 652
 653 int fio_monotonic_clocktest(int debug)
 654 {
 655         struct clock_thread *cthreads;
 656         unsigned int nr_cpus = cpus_online();
 657         struct clock_entry *entries;
 658         unsigned long nr_entries, tentries, failed = 0;
 659         struct clock_entry *prev, *this;
 660         uint32_t seq = 0;
 661         unsigned int i;
 662
 663         if (debug) {
 664                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 665
 666 #ifdef FIO_INC_DEBUG
 667                 fio_debug |= 1U << FD_TIME;
 668 #endif
 669                 nr_entries = CLOCK_ENTRIES_DEBUG;
 670         } else
 671                 nr_entries = CLOCK_ENTRIES_TEST;
 672
 673         calibrate_cpu_clock();
 674
 675         if (debug) {
 676 #ifdef FIO_INC_DEBUG
 677                 fio_debug &= ~(1U << FD_TIME);
 678 #endif
 679         }
 680
 681         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 682         tentries = nr_entries * nr_cpus;
 683         entries = malloc(tentries * sizeof(struct clock_entry));
 684
 685         if (debug)
 686                 log_info("cs: Testing %u CPUs\n", nr_cpus);
 687
 688         for (i = 0; i < nr_cpus; i++) {
 689                 struct clock_thread *t = &cthreads[i];
 690
 691                 t->cpu = i;
 692                 t->debug = debug;
 693                 t->seq = &seq;
 694                 t->nr_entries = nr_entries;
 695                 t->entries = &entries[i * nr_entries];
 696                 __fio_sem_init(&t->lock, FIO_SEM_LOCKED);
 697                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 698                         failed++;
 699                         nr_cpus = i;
 700                         break;
 701                 }
 702         }
 703
 704         for (i = 0; i < nr_cpus; i++) {
 705                 struct clock_thread *t = &cthreads[i];
 706
 707                 fio_sem_up(&t->lock);
 708         }
 709
 710         for (i = 0; i < nr_cpus; i++) {
 711                 struct clock_thread *t = &cthreads[i];
 712                 void *ret;
 713
 714                 pthread_join(t->thread, &ret);
 715                 if (ret)
 716                         failed++;
 717                 __fio_sem_remove(&t->lock);
 718         }
 719         free(cthreads);
 720
 721         if (failed) {
 722                 if (debug)
 723                         log_err("Clocksource test: %lu threads failed\n", failed);
 724                 goto err;
 725         }
 726
 727         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 728
 729         /* silence silly gcc */
 730         prev = NULL;
 731         for (failed = i = 0; i < tentries; i++) {
 732                 this = &entries[i];
 733
 734                 if (!i) {
 735                         prev = this;
 736                         continue;
 737                 }
 738
 739                 if (prev->tsc > this->tsc) {
 740                         uint64_t diff = prev->tsc - this->tsc;
 741
 742                         if (!debug) {
 743                                 failed++;
 744                                 break;
 745                         }
 746
 747                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 748                                                 (unsigned long long) diff);
 749                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 750                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 751                         failed++;
 752                 }
 753
 754                 prev = this;
 755         }
 756
 757         if (debug) {
 758                 if (failed)
 759                         log_info("cs: Failed: %lu\n", failed);
 760                 else
 761                         log_info("cs: Pass!\n");
 762         }
 763 err:
 764         free(entries);
 765         return !!failed;
 766 }
 767
 768 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 769
 770 int fio_monotonic_clocktest(int debug)
 771 {
 772         if (debug)
 773                 log_info("cs: current platform does not support CPU clocks\n");
 774         return 1;
 775 }
 776
 777 #endif