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)
  17 #ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
  18 static unsigned long long cycles_per_msec;
  19 static unsigned long long cycles_start;
  20 static unsigned long long clock_mult;
  21 static unsigned long long max_cycles_mask;
  22 static unsigned long long nsecs_for_max_cycles;
  23 static unsigned int clock_shift;
  24 static unsigned int max_cycles_shift;
  25 #define MAX_CLOCK_SEC 60*60
  26 #endif
  27 #ifdef ARCH_CPU_CLOCK_WRAPS
  28 static unsigned int cycles_wrap;
  29 #endif
  30 #endif
  31 bool tsc_reliable = false;
  32
  33 struct tv_valid {
  34         int warned;
  35 };
  36 #ifdef ARCH_HAVE_CPU_CLOCK
  37 #ifdef CONFIG_TLS_THREAD
  38 static __thread struct tv_valid static_tv_valid;
  39 #else
  40 static pthread_key_t tv_tls_key;
  41 #endif
  42 #endif
  43
  44 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  45 int fio_clock_source_set = 0;
  46 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  47
  48 #ifdef FIO_DEBUG_TIME
  49
  50 #define HASH_BITS       8
  51 #define HASH_SIZE       (1 << HASH_BITS)
  52
  53 static struct flist_head hash[HASH_SIZE];
  54 static int gtod_inited;
  55
  56 struct gtod_log {
  57         struct flist_head list;
  58         void *caller;
  59         unsigned long calls;
  60 };
  61
  62 static struct gtod_log *find_hash(void *caller)
  63 {
  64         unsigned long h = hash_ptr(caller, HASH_BITS);
  65         struct flist_head *entry;
  66
  67         flist_for_each(entry, &hash[h]) {
  68                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  69                                                                         list);
  70
  71                 if (log->caller == caller)
  72                         return log;
  73         }
  74
  75         return NULL;
  76 }
  77
  78 static void inc_caller(void *caller)
  79 {
  80         struct gtod_log *log = find_hash(caller);
  81
  82         if (!log) {
  83                 unsigned long h;
  84
  85                 log = malloc(sizeof(*log));
  86                 INIT_FLIST_HEAD(&log->list);
  87                 log->caller = caller;
  88                 log->calls = 0;
  89
  90                 h = hash_ptr(caller, HASH_BITS);
  91                 flist_add_tail(&log->list, &hash[h]);
  92         }
  93
  94         log->calls++;
  95 }
  96
  97 static void gtod_log_caller(void *caller)
  98 {
  99         if (gtod_inited)
 100                 inc_caller(caller);
 101 }
 102
 103 static void fio_exit fio_dump_gtod(void)
 104 {
 105         unsigned long total_calls = 0;
 106         int i;
 107
 108         for (i = 0; i < HASH_SIZE; i++) {
 109                 struct flist_head *entry;
 110                 struct gtod_log *log;
 111
 112                 flist_for_each(entry, &hash[i]) {
 113                         log = flist_entry(entry, struct gtod_log, list);
 114
 115                         printf("function %p, calls %lu\n", log->caller,
 116                                                                 log->calls);
 117                         total_calls += log->calls;
 118                 }
 119         }
 120
 121         printf("Total %lu gettimeofday\n", total_calls);
 122 }
 123
 124 static void fio_init gtod_init(void)
 125 {
 126         int i;
 127
 128         for (i = 0; i < HASH_SIZE; i++)
 129                 INIT_FLIST_HEAD(&hash[i]);
 130
 131         gtod_inited = 1;
 132 }
 133
 134 #endif /* FIO_DEBUG_TIME */
 135
 136 #ifdef CONFIG_CLOCK_GETTIME
 137 static int fill_clock_gettime(struct timespec *ts)
 138 {
 139 #if defined(CONFIG_CLOCK_MONOTONIC_RAW)
 140         return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
 141 #elif defined(CONFIG_CLOCK_MONOTONIC)
 142         return clock_gettime(CLOCK_MONOTONIC, ts);
 143 #else
 144         return clock_gettime(CLOCK_REALTIME, ts);
 145 #endif
 146 }
 147 #endif
 148
 149 static void __fio_gettime(struct timespec *tp)
 150 {
 151         switch (fio_clock_source) {
 152 #ifdef CONFIG_GETTIMEOFDAY
 153         case CS_GTOD: {
 154                 struct timeval tv;
 155                 gettimeofday(&tv, NULL);
 156
 157                 tp->tv_sec = tv.tv_sec;
 158                 tp->tv_nsec = tv.tv_usec * 1000;
 159                 break;
 160                 }
 161 #endif
 162 #ifdef CONFIG_CLOCK_GETTIME
 163         case CS_CGETTIME: {
 164                 if (fill_clock_gettime(tp) < 0) {
 165                         log_err("fio: clock_gettime fails\n");
 166                         assert(0);
 167                 }
 168                 break;
 169                 }
 170 #endif
 171 #ifdef ARCH_HAVE_CPU_CLOCK
 172         case CS_CPUCLOCK: {
 173                 uint64_t nsecs, t, multiples;
 174                 struct tv_valid *tv;
 175
 176 #ifdef CONFIG_TLS_THREAD
 177                 tv = &static_tv_valid;
 178 #else
 179                 tv = pthread_getspecific(tv_tls_key);
 180 #endif
 181
 182                 t = get_cpu_clock();
 183 #ifdef ARCH_CPU_CLOCK_WRAPS
 184                 if (t < cycles_start && !cycles_wrap)
 185                         cycles_wrap = 1;
 186                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
 187                         log_err("fio: double CPU clock wrap\n");
 188                         tv->warned = 1;
 189                 }
 190 #endif
 191 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 192                 nsecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC * 1000;
 193 #else
 194                 t -= cycles_start;
 195                 multiples = t >> max_cycles_shift;
 196                 nsecs = multiples * nsecs_for_max_cycles;
 197                 nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift;
 198 #endif
 199                 tp->tv_sec = nsecs / 1000000000ULL;
 200                 tp->tv_nsec = nsecs % 1000000000ULL;
 201                 break;
 202                 }
 203 #endif
 204         default:
 205                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 206                 break;
 207         }
 208 }
 209
 210 #ifdef FIO_DEBUG_TIME
 211 void fio_gettime(struct timespec *tp, void *caller)
 212 #else
 213 void fio_gettime(struct timespec *tp, void fio_unused *caller)
 214 #endif
 215 {
 216 #ifdef FIO_DEBUG_TIME
 217         if (!caller)
 218                 caller = __builtin_return_address(0);
 219
 220         gtod_log_caller(caller);
 221 #endif
 222         if (fio_unlikely(fio_gettime_offload(tp)))
 223                 return;
 224
 225         __fio_gettime(tp);
 226 }
 227
 228 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 229 static unsigned long get_cycles_per_msec(void)
 230 {
 231         struct timespec s, e;
 232         uint64_t c_s, c_e;
 233         enum fio_cs old_cs = fio_clock_source;
 234         uint64_t elapsed;
 235
 236 #ifdef CONFIG_CLOCK_GETTIME
 237         fio_clock_source = CS_CGETTIME;
 238 #else
 239         fio_clock_source = CS_GTOD;
 240 #endif
 241         __fio_gettime(&s);
 242
 243         c_s = get_cpu_clock();
 244         do {
 245                 __fio_gettime(&e);
 246
 247                 elapsed = utime_since(&s, &e);
 248                 if (elapsed >= 1280) {
 249                         c_e = get_cpu_clock();
 250                         break;
 251                 }
 252         } while (1);
 253
 254         fio_clock_source = old_cs;
 255         return (c_e - c_s) * 1000 / elapsed;
 256 }
 257
 258 #define NR_TIME_ITERS   50
 259
 260 static int calibrate_cpu_clock(void)
 261 {
 262         double delta, mean, S;
 263         uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
 264         int i, samples, sft = 0;
 265         unsigned long long tmp, max_ticks, max_mult;
 266
 267         cycles[0] = get_cycles_per_msec();
 268         S = delta = mean = 0.0;
 269         for (i = 0; i < NR_TIME_ITERS; i++) {
 270                 cycles[i] = get_cycles_per_msec();
 271                 delta = cycles[i] - mean;
 272                 if (delta) {
 273                         mean += delta / (i + 1.0);
 274                         S += delta * (cycles[i] - mean);
 275                 }
 276         }
 277
 278         /*
 279          * The most common platform clock breakage is returning zero
 280          * indefinitely. Check for that and return failure.
 281          */
 282         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 283                 return 1;
 284
 285         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 286
 287         minc = -1ULL;
 288         maxc = samples = avg = 0;
 289         for (i = 0; i < NR_TIME_ITERS; i++) {
 290                 double this = cycles[i];
 291
 292                 minc = min(cycles[i], minc);
 293                 maxc = max(cycles[i], maxc);
 294
 295                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 296                         continue;
 297                 samples++;
 298                 avg += this;
 299         }
 300
 301         S /= (double) NR_TIME_ITERS;
 302
 303         for (i = 0; i < NR_TIME_ITERS; i++)
 304                 dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
 305
 306         avg /= samples;
 307         cycles_per_msec = avg;
 308         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 309         dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
 310                         (unsigned long long) minc,
 311                         (unsigned long long) maxc, mean, S);
 312
 313         max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL;
 314         max_mult = ULLONG_MAX / max_ticks;
 315         dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, "
 316                         "max_mult=%llu\n", max_ticks,
 317                         __builtin_clzll(max_ticks), max_mult);
 318
 319         /*
 320          * Find the largest shift count that will produce
 321          * a multiplier that does not exceed max_mult
 322          */
 323         tmp = max_mult * cycles_per_msec / 1000000;
 324         while (tmp > 1) {
 325                 tmp >>= 1;
 326                 sft++;
 327                 dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft);
 328         }
 329
 330         clock_shift = sft;
 331         clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec;
 332         dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift,
 333                                                         clock_mult);
 334
 335         /*
 336          * Find the greatest power of 2 clock ticks that is less than the
 337          * ticks in MAX_CLOCK_SEC_2STAGE
 338          */
 339         max_cycles_shift = max_cycles_mask = 0;
 340         tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec;
 341         dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp,
 342                                                         max_cycles_shift);
 343         while (tmp > 1) {
 344                 tmp >>= 1;
 345                 max_cycles_shift++;
 346                 dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
 347         }
 348         /*
 349          * if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec
 350          * here we will have a discontinuity every
 351          * (1ULL << max_cycles_shift) cycles
 352          */
 353         nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult)
 354                                         >> clock_shift;
 355
 356         /* Use a bitmask to calculate ticks % (1ULL << max_cycles_shift) */
 357         for (tmp = 0; tmp < max_cycles_shift; tmp++)
 358                 max_cycles_mask |= 1ULL << tmp;
 359
 360         dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, "
 361                         "nsecs_for_max_cycles=%llu, "
 362                         "max_cycles_mask=%016llx\n",
 363                         max_cycles_shift, (1ULL << max_cycles_shift),
 364                         nsecs_for_max_cycles, max_cycles_mask);
 365
 366         cycles_start = get_cpu_clock();
 367         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
 368         return 0;
 369 }
 370 #else
 371 static int calibrate_cpu_clock(void)
 372 {
 373 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 374         return 0;
 375 #else
 376         return 1;
 377 #endif
 378 }
 379 #endif // ARCH_HAVE_CPU_CLOCK
 380
 381 #ifndef CONFIG_TLS_THREAD
 382 void fio_local_clock_init(int is_thread)
 383 {
 384         struct tv_valid *t;
 385
 386         t = calloc(1, sizeof(*t));
 387         if (pthread_setspecific(tv_tls_key, t)) {
 388                 log_err("fio: can't set TLS key\n");
 389                 assert(0);
 390         }
 391 }
 392
 393 static void kill_tv_tls_key(void *data)
 394 {
 395         free(data);
 396 }
 397 #else
 398 void fio_local_clock_init(int is_thread)
 399 {
 400 }
 401 #endif
 402
 403 void fio_clock_init(void)
 404 {
 405         if (fio_clock_source == fio_clock_source_inited)
 406                 return;
 407
 408 #ifndef CONFIG_TLS_THREAD
 409         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 410                 log_err("fio: can't create TLS key\n");
 411 #endif
 412
 413         fio_clock_source_inited = fio_clock_source;
 414
 415         if (calibrate_cpu_clock())
 416                 tsc_reliable = false;
 417
 418         /*
 419          * If the arch sets tsc_reliable != 0, then it must be good enough
 420          * to use as THE clock source. For x86 CPUs, this means the TSC
 421          * runs at a constant rate and is synced across CPU cores.
 422          */
 423         if (tsc_reliable) {
 424                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
 425                         fio_clock_source = CS_CPUCLOCK;
 426         } else if (fio_clock_source == CS_CPUCLOCK)
 427                 log_info("fio: clocksource=cpu may not be reliable\n");
 428         dprint(FD_TIME, "gettime: clocksource=%d\n", (int) fio_clock_source);
 429 }
 430
 431 uint64_t ntime_since(const struct timespec *s, const struct timespec *e)
 432 {
 433        int64_t sec, nsec;
 434
 435        sec = e->tv_sec - s->tv_sec;
 436        nsec = e->tv_nsec - s->tv_nsec;
 437        if (sec > 0 && nsec < 0) {
 438                sec--;
 439                nsec += 1000000000LL;
 440        }
 441
 442        /*
 443         * time warp bug on some kernels?
 444         */
 445        if (sec < 0 || (sec == 0 && nsec < 0))
 446                return 0;
 447
 448        return nsec + (sec * 1000000000LL);
 449 }
 450
 451 uint64_t ntime_since_now(const struct timespec *s)
 452 {
 453         struct timespec now;
 454
 455         fio_gettime(&now, NULL);
 456         return ntime_since(s, &now);
 457 }
 458
 459 uint64_t utime_since(const struct timespec *s, const struct timespec *e)
 460 {
 461         int64_t sec, usec;
 462
 463         sec = e->tv_sec - s->tv_sec;
 464         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 465         if (sec > 0 && usec < 0) {
 466                 sec--;
 467                 usec += 1000000;
 468         }
 469
 470         /*
 471          * time warp bug on some kernels?
 472          */
 473         if (sec < 0 || (sec == 0 && usec < 0))
 474                 return 0;
 475
 476         return usec + (sec * 1000000);
 477 }
 478
 479 uint64_t utime_since_now(const struct timespec *s)
 480 {
 481         struct timespec t;
 482 #ifdef FIO_DEBUG_TIME
 483         void *p = __builtin_return_address(0);
 484
 485         fio_gettime(&t, p);
 486 #else
 487         fio_gettime(&t, NULL);
 488 #endif
 489
 490         return utime_since(s, &t);
 491 }
 492
 493 uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
 494 {
 495         int64_t sec, usec;
 496
 497         sec = e->tv_sec - s->tv_sec;
 498         usec = (e->tv_usec - s->tv_usec);
 499         if (sec > 0 && usec < 0) {
 500                 sec--;
 501                 usec += 1000000;
 502         }
 503
 504         if (sec < 0 || (sec == 0 && usec < 0))
 505                 return 0;
 506
 507         sec *= 1000;
 508         usec /= 1000;
 509         return sec + usec;
 510 }
 511
 512 uint64_t mtime_since_now(const struct timespec *s)
 513 {
 514         struct timespec t;
 515 #ifdef FIO_DEBUG_TIME
 516         void *p = __builtin_return_address(0);
 517
 518         fio_gettime(&t, p);
 519 #else
 520         fio_gettime(&t, NULL);
 521 #endif
 522
 523         return mtime_since(s, &t);
 524 }
 525
 526 uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
 527 {
 528         int64_t sec, usec;
 529
 530         sec = e->tv_sec - s->tv_sec;
 531         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 532         if (sec > 0 && usec < 0) {
 533                 sec--;
 534                 usec += 1000000;
 535         }
 536
 537         if (sec < 0 || (sec == 0 && usec < 0))
 538                 return 0;
 539
 540         sec *= 1000;
 541         usec /= 1000;
 542         return sec + usec;
 543 }
 544
 545 uint64_t time_since_now(const struct timespec *s)
 546 {
 547         return mtime_since_now(s) / 1000;
 548 }
 549
 550 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 551     defined(CONFIG_SYNC_SYNC) && defined(CONFIG_CMP_SWAP)
 552
 553 #define CLOCK_ENTRIES_DEBUG     100000
 554 #define CLOCK_ENTRIES_TEST      1000
 555
 556 struct clock_entry {
 557         uint32_t seq;
 558         uint32_t cpu;
 559         uint64_t tsc;
 560 };
 561
 562 struct clock_thread {
 563         pthread_t thread;
 564         int cpu;
 565         int debug;
 566         pthread_mutex_t lock;
 567         pthread_mutex_t started;
 568         unsigned long nr_entries;
 569         uint32_t *seq;
 570         struct clock_entry *entries;
 571 };
 572
 573 static inline uint32_t atomic32_compare_and_swap(uint32_t *ptr, uint32_t old,
 574                                                  uint32_t new)
 575 {
 576         return __sync_val_compare_and_swap(ptr, old, new);
 577 }
 578
 579 static void *clock_thread_fn(void *data)
 580 {
 581         struct clock_thread *t = data;
 582         struct clock_entry *c;
 583         os_cpu_mask_t cpu_mask;
 584         unsigned long long first;
 585         int i;
 586
 587         if (fio_cpuset_init(&cpu_mask)) {
 588                 int __err = errno;
 589
 590                 log_err("clock cpuset init failed: %s\n", strerror(__err));
 591                 goto err_out;
 592         }
 593
 594         fio_cpu_set(&cpu_mask, t->cpu);
 595
 596         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 597                 int __err = errno;
 598
 599                 log_err("clock setaffinity failed: %s\n", strerror(__err));
 600                 goto err;
 601         }
 602
 603         pthread_mutex_lock(&t->lock);
 604         pthread_mutex_unlock(&t->started);
 605
 606         first = get_cpu_clock();
 607         c = &t->entries[0];
 608         for (i = 0; i < t->nr_entries; i++, c++) {
 609                 uint32_t seq;
 610                 uint64_t tsc;
 611
 612                 c->cpu = t->cpu;
 613                 do {
 614                         seq = *t->seq;
 615                         if (seq == UINT_MAX)
 616                                 break;
 617                         __sync_synchronize();
 618                         tsc = get_cpu_clock();
 619                 } while (seq != atomic32_compare_and_swap(t->seq, seq, seq + 1));
 620
 621                 if (seq == UINT_MAX)
 622                         break;
 623
 624                 c->seq = seq;
 625                 c->tsc = tsc;
 626         }
 627
 628         if (t->debug) {
 629                 unsigned long long clocks;
 630
 631                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
 632                 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
 633                                                         clocks, first);
 634         }
 635
 636         /*
 637          * The most common platform clock breakage is returning zero
 638          * indefinitely. Check for that and return failure.
 639          */
 640         if (i > 1 && !t->entries[i - 1].tsc && !t->entries[0].tsc)
 641                 goto err;
 642
 643         fio_cpuset_exit(&cpu_mask);
 644         return NULL;
 645 err:
 646         fio_cpuset_exit(&cpu_mask);
 647 err_out:
 648         return (void *) 1;
 649 }
 650
 651 static int clock_cmp(const void *p1, const void *p2)
 652 {
 653         const struct clock_entry *c1 = p1;
 654         const struct clock_entry *c2 = p2;
 655
 656         if (c1->seq == c2->seq)
 657                 log_err("cs: bug in atomic sequence!\n");
 658
 659         return c1->seq - c2->seq;
 660 }
 661
 662 int fio_monotonic_clocktest(int debug)
 663 {
 664         struct clock_thread *cthreads;
 665         unsigned int nr_cpus = cpus_online();
 666         struct clock_entry *entries;
 667         unsigned long nr_entries, tentries, failed = 0;
 668         struct clock_entry *prev, *this;
 669         uint32_t seq = 0;
 670         unsigned int i;
 671
 672         if (debug) {
 673                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 674
 675 #ifdef FIO_INC_DEBUG
 676                 fio_debug |= 1U << FD_TIME;
 677 #endif
 678                 nr_entries = CLOCK_ENTRIES_DEBUG;
 679         } else
 680                 nr_entries = CLOCK_ENTRIES_TEST;
 681
 682         calibrate_cpu_clock();
 683
 684         if (debug) {
 685 #ifdef FIO_INC_DEBUG
 686                 fio_debug &= ~(1U << FD_TIME);
 687 #endif
 688         }
 689
 690         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 691         tentries = nr_entries * nr_cpus;
 692         entries = malloc(tentries * sizeof(struct clock_entry));
 693
 694         if (debug)
 695                 log_info("cs: Testing %u CPUs\n", nr_cpus);
 696
 697         for (i = 0; i < nr_cpus; i++) {
 698                 struct clock_thread *t = &cthreads[i];
 699
 700                 t->cpu = i;
 701                 t->debug = debug;
 702                 t->seq = &seq;
 703                 t->nr_entries = nr_entries;
 704                 t->entries = &entries[i * nr_entries];
 705                 pthread_mutex_init(&t->lock, NULL);
 706                 pthread_mutex_init(&t->started, NULL);
 707                 pthread_mutex_lock(&t->lock);
 708                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 709                         failed++;
 710                         nr_cpus = i;
 711                         break;
 712                 }
 713         }
 714
 715         for (i = 0; i < nr_cpus; i++) {
 716                 struct clock_thread *t = &cthreads[i];
 717
 718                 pthread_mutex_lock(&t->started);
 719         }
 720
 721         for (i = 0; i < nr_cpus; i++) {
 722                 struct clock_thread *t = &cthreads[i];
 723
 724                 pthread_mutex_unlock(&t->lock);
 725         }
 726
 727         for (i = 0; i < nr_cpus; i++) {
 728                 struct clock_thread *t = &cthreads[i];
 729                 void *ret;
 730
 731                 pthread_join(t->thread, &ret);
 732                 if (ret)
 733                         failed++;
 734         }
 735         free(cthreads);
 736
 737         if (failed) {
 738                 if (debug)
 739                         log_err("Clocksource test: %lu threads failed\n", failed);
 740                 goto err;
 741         }
 742
 743         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 744
 745         /* silence silly gcc */
 746         prev = NULL;
 747         for (failed = i = 0; i < tentries; i++) {
 748                 this = &entries[i];
 749
 750                 if (!i) {
 751                         prev = this;
 752                         continue;
 753                 }
 754
 755                 if (prev->tsc > this->tsc) {
 756                         uint64_t diff = prev->tsc - this->tsc;
 757
 758                         if (!debug) {
 759                                 failed++;
 760                                 break;
 761                         }
 762
 763                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 764                                                 (unsigned long long) diff);
 765                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 766                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 767                         failed++;
 768                 }
 769
 770                 prev = this;
 771         }
 772
 773         if (debug) {
 774                 if (failed)
 775                         log_info("cs: Failed: %lu\n", failed);
 776                 else
 777                         log_info("cs: Pass!\n");
 778         }
 779 err:
 780         free(entries);
 781         return !!failed;
 782 }
 783
 784 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 785
 786 int fio_monotonic_clocktest(int debug)
 787 {
 788         if (debug)
 789                 log_info("cs: current platform does not support CPU clocks\n");
 790         return 1;
 791 }
 792
 793 #endif