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 cycles_per_usec;
  19 static unsigned long inv_cycles_per_nsec;
  20 static uint64_t max_cycles_for_mult;
  21 #define NSEC_INV_FACTOR 4096
  22 #endif
  23 #ifdef ARCH_CPU_CLOCK_WRAPS
  24 static unsigned long long cycles_start, cycles_wrap;
  25 #endif
  26 #endif
  27 int tsc_reliable = 0;
  28
  29 struct tv_valid {
  30         uint64_t last_cycles;
  31         int last_tv_valid;
  32         int warned;
  33 };
  34 #ifdef ARCH_HAVE_CPU_CLOCK
  35 #ifdef CONFIG_TLS_THREAD
  36 static __thread struct tv_valid static_tv_valid;
  37 #else
  38 static pthread_key_t tv_tls_key;
  39 #endif
  40 #endif
  41
  42 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  43 int fio_clock_source_set = 0;
  44 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  45
  46 #ifdef FIO_DEBUG_TIME
  47
  48 #define HASH_BITS       8
  49 #define HASH_SIZE       (1 << HASH_BITS)
  50
  51 static struct flist_head hash[HASH_SIZE];
  52 static int gtod_inited;
  53
  54 struct gtod_log {
  55         struct flist_head list;
  56         void *caller;
  57         unsigned long calls;
  58 };
  59
  60 static struct gtod_log *find_hash(void *caller)
  61 {
  62         unsigned long h = hash_ptr(caller, HASH_BITS);
  63         struct flist_head *entry;
  64
  65         flist_for_each(entry, &hash[h]) {
  66                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  67                                                                         list);
  68
  69                 if (log->caller == caller)
  70                         return log;
  71         }
  72
  73         return NULL;
  74 }
  75
  76 static void inc_caller(void *caller)
  77 {
  78         struct gtod_log *log = find_hash(caller);
  79
  80         if (!log) {
  81                 unsigned long h;
  82
  83                 log = malloc(sizeof(*log));
  84                 INIT_FLIST_HEAD(&log->list);
  85                 log->caller = caller;
  86                 log->calls = 0;
  87
  88                 h = hash_ptr(caller, HASH_BITS);
  89                 flist_add_tail(&log->list, &hash[h]);
  90         }
  91
  92         log->calls++;
  93 }
  94
  95 static void gtod_log_caller(void *caller)
  96 {
  97         if (gtod_inited)
  98                 inc_caller(caller);
  99 }
 100
 101 static void fio_exit fio_dump_gtod(void)
 102 {
 103         unsigned long total_calls = 0;
 104         int i;
 105
 106         for (i = 0; i < HASH_SIZE; i++) {
 107                 struct flist_head *entry;
 108                 struct gtod_log *log;
 109
 110                 flist_for_each(entry, &hash[i]) {
 111                         log = flist_entry(entry, struct gtod_log, list);
 112
 113                         printf("function %p, calls %lu\n", log->caller,
 114                                                                 log->calls);
 115                         total_calls += log->calls;
 116                 }
 117         }
 118
 119         printf("Total %lu gettimeofday\n", total_calls);
 120 }
 121
 122 static void fio_init gtod_init(void)
 123 {
 124         int i;
 125
 126         for (i = 0; i < HASH_SIZE; i++)
 127                 INIT_FLIST_HEAD(&hash[i]);
 128
 129         gtod_inited = 1;
 130 }
 131
 132 #endif /* FIO_DEBUG_TIME */
 133
 134 #ifdef CONFIG_CLOCK_GETTIME
 135 static int fill_clock_gettime(struct timespec *ts)
 136 {
 137 #if defined(CONFIG_CLOCK_MONOTONIC_RAW)
 138         return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
 139 #elif defined(CONFIG_CLOCK_MONOTONIC)
 140         return clock_gettime(CLOCK_MONOTONIC, ts);
 141 #else
 142         return clock_gettime(CLOCK_REALTIME, ts);
 143 #endif
 144 }
 145 #endif
 146
 147 static void __fio_gettime(struct timespec *tp)
 148 {
 149         switch (fio_clock_source) {
 150 #ifdef CONFIG_GETTIMEOFDAY
 151         case CS_GTOD: {
 152                 struct timeval tv;
 153                 gettimeofday(&tv, NULL);
 154
 155                 tp->tv_sec = tv.tv_sec;
 156                 tp->tv_nsec = tv.tv_usec * 1000;
 157                 break;
 158                 }
 159 #endif
 160 #ifdef CONFIG_CLOCK_GETTIME
 161         case CS_CGETTIME: {
 162                 if (fill_clock_gettime(tp) < 0) {
 163                         log_err("fio: clock_gettime fails\n");
 164                         assert(0);
 165                 }
 166                 break;
 167                 }
 168 #endif
 169 #ifdef ARCH_HAVE_CPU_CLOCK
 170         case CS_CPUCLOCK: {
 171                 uint64_t nsecs, t;
 172                 struct tv_valid *tv;
 173
 174 #ifdef CONFIG_TLS_THREAD
 175                 tv = &static_tv_valid;
 176 #else
 177                 tv = pthread_getspecific(tv_tls_key);
 178 #endif
 179
 180                 t = get_cpu_clock();
 181 #ifdef ARCH_CPU_CLOCK_WRAPS
 182                 if (t < cycles_start && !cycles_wrap)
 183                         cycles_wrap = 1;
 184                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
 185                         log_err("fio: double CPU clock wrap\n");
 186                         tv->warned = 1;
 187                 }
 188
 189                 t -= cycles_start;
 190 #endif
 191                 tv->last_cycles = t;
 192                 tv->last_tv_valid = 1;
 193 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 194                 nsecs = t * 1000 / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 195 #else
 196                 if (t < max_cycles_for_mult)
 197                         nsecs = (t * inv_cycles_per_nsec) / NSEC_INV_FACTOR;
 198                 else
 199                         nsecs = (t / NSEC_INV_FACTOR) * inv_cycles_per_nsec;
 200 #endif
 201                 tp->tv_sec = nsecs / 1000000000ULL;
 202                 tp->tv_nsec = nsecs % 1000000000ULL;
 203                 break;
 204                 }
 205 #endif
 206         default:
 207                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 208                 break;
 209         }
 210 }
 211
 212 #ifdef FIO_DEBUG_TIME
 213 void fio_gettime(struct timespec *tp, void *caller)
 214 #else
 215 void fio_gettime(struct timespec *tp, void fio_unused *caller)
 216 #endif
 217 {
 218 #ifdef FIO_DEBUG_TIME
 219         if (!caller)
 220                 caller = __builtin_return_address(0);
 221
 222         gtod_log_caller(caller);
 223 #endif
 224         if (fio_unlikely(fio_gettime_offload(tp)))
 225                 return;
 226
 227         __fio_gettime(tp);
 228 }
 229
 230 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 231 static unsigned long get_cycles_per_usec(void)
 232 {
 233         struct timespec s, e;
 234         uint64_t c_s, c_e;
 235         enum fio_cs old_cs = fio_clock_source;
 236         uint64_t elapsed;
 237
 238 #ifdef CONFIG_CLOCK_GETTIME
 239         fio_clock_source = CS_CGETTIME;
 240 #else
 241         fio_clock_source = CS_GTOD;
 242 #endif
 243         __fio_gettime(&s);
 244
 245         c_s = get_cpu_clock();
 246         do {
 247                 __fio_gettime(&e);
 248
 249                 elapsed = utime_since(&s, &e);
 250                 if (elapsed >= 1280) {
 251                         c_e = get_cpu_clock();
 252                         break;
 253                 }
 254         } while (1);
 255
 256         fio_clock_source = old_cs;
 257         return (c_e - c_s) / elapsed;
 258 }
 259
 260 #define NR_TIME_ITERS   50
 261
 262 static int calibrate_cpu_clock(void)
 263 {
 264         double delta, mean, S;
 265         uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
 266         int i, samples;
 267
 268         cycles[0] = get_cycles_per_usec();
 269         S = delta = mean = 0.0;
 270         for (i = 0; i < NR_TIME_ITERS; i++) {
 271                 cycles[i] = get_cycles_per_usec();
 272                 delta = cycles[i] - mean;
 273                 if (delta) {
 274                         mean += delta / (i + 1.0);
 275                         S += delta * (cycles[i] - mean);
 276                 }
 277         }
 278
 279         /*
 280          * The most common platform clock breakage is returning zero
 281          * indefinitely. Check for that and return failure.
 282          */
 283         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 284                 return 1;
 285
 286         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 287
 288         minc = -1ULL;
 289         maxc = samples = avg = 0;
 290         for (i = 0; i < NR_TIME_ITERS; i++) {
 291                 double this = cycles[i];
 292
 293                 minc = min(cycles[i], minc);
 294                 maxc = max(cycles[i], maxc);
 295
 296                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 297                         continue;
 298                 samples++;
 299                 avg += this;
 300         }
 301
 302         S /= (double) NR_TIME_ITERS;
 303
 304         for (i = 0; i < NR_TIME_ITERS; i++)
 305                 dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
 306
 307         avg /= samples;
 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         cycles_per_usec = avg;
 314         inv_cycles_per_nsec = NSEC_INV_FACTOR * 1000 / cycles_per_usec;
 315         max_cycles_for_mult = ~0ULL / inv_cycles_per_nsec;
 316         dprint(FD_TIME, "inv_cycles_per_nsec=%lu\n", inv_cycles_per_nsec);
 317 #ifdef ARCH_CPU_CLOCK_WRAPS
 318         cycles_start = get_cpu_clock();
 319         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
 320 #endif
 321         return 0;
 322 }
 323 #else
 324 static int calibrate_cpu_clock(void)
 325 {
 326 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 327         return 0;
 328 #else
 329         return 1;
 330 #endif
 331 }
 332 #endif // ARCH_HAVE_CPU_CLOCK
 333
 334 #ifndef CONFIG_TLS_THREAD
 335 void fio_local_clock_init(int is_thread)
 336 {
 337         struct tv_valid *t;
 338
 339         t = calloc(1, sizeof(*t));
 340         if (pthread_setspecific(tv_tls_key, t)) {
 341                 log_err("fio: can't set TLS key\n");
 342                 assert(0);
 343         }
 344 }
 345
 346 static void kill_tv_tls_key(void *data)
 347 {
 348         free(data);
 349 }
 350 #else
 351 void fio_local_clock_init(int is_thread)
 352 {
 353 }
 354 #endif
 355
 356 void fio_clock_init(void)
 357 {
 358         if (fio_clock_source == fio_clock_source_inited)
 359                 return;
 360
 361 #ifndef CONFIG_TLS_THREAD
 362         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 363                 log_err("fio: can't create TLS key\n");
 364 #endif
 365
 366         fio_clock_source_inited = fio_clock_source;
 367
 368         if (calibrate_cpu_clock())
 369                 tsc_reliable = 0;
 370
 371         /*
 372          * If the arch sets tsc_reliable != 0, then it must be good enough
 373          * to use as THE clock source. For x86 CPUs, this means the TSC
 374          * runs at a constant rate and is synced across CPU cores.
 375          */
 376         if (tsc_reliable) {
 377                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
 378                         fio_clock_source = CS_CPUCLOCK;
 379         } else if (fio_clock_source == CS_CPUCLOCK)
 380                 log_info("fio: clocksource=cpu may not be reliable\n");
 381 }
 382
 383 uint64_t utime_since(const struct timespec *s, const struct timespec *e)
 384 {
 385         int64_t sec, usec;
 386
 387         sec = e->tv_sec - s->tv_sec;
 388         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 389         if (sec > 0 && usec < 0) {
 390                 sec--;
 391                 usec += 1000000;
 392         }
 393
 394         /*
 395          * time warp bug on some kernels?
 396          */
 397         if (sec < 0 || (sec == 0 && usec < 0))
 398                 return 0;
 399
 400         return usec + (sec * 1000000);
 401 }
 402
 403 uint64_t utime_since_now(const struct timespec *s)
 404 {
 405         struct timespec t;
 406 #ifdef FIO_DEBUG_TIME
 407         void *p = __builtin_return_address(0);
 408
 409         fio_gettime(&t, p);
 410 #else
 411         fio_gettime(&t, NULL);
 412 #endif
 413
 414         return utime_since(s, &t);
 415 }
 416
 417 uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
 418 {
 419         int64_t sec, usec;
 420
 421         sec = e->tv_sec - s->tv_sec;
 422         usec = (e->tv_usec - s->tv_usec);
 423         if (sec > 0 && usec < 0) {
 424                 sec--;
 425                 usec += 1000000;
 426         }
 427
 428         if (sec < 0 || (sec == 0 && usec < 0))
 429                 return 0;
 430
 431         sec *= 1000;
 432         usec /= 1000;
 433         return sec + usec;
 434 }
 435
 436 uint64_t mtime_since_now(const struct timespec *s)
 437 {
 438         struct timespec t;
 439 #ifdef FIO_DEBUG_TIME
 440         void *p = __builtin_return_address(0);
 441
 442         fio_gettime(&t, p);
 443 #else
 444         fio_gettime(&t, NULL);
 445 #endif
 446
 447         return mtime_since(s, &t);
 448 }
 449
 450 uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
 451 {
 452         int64_t sec, usec;
 453
 454         sec = e->tv_sec - s->tv_sec;
 455         usec = (e->tv_nsec - s->tv_nsec) / 1000;
 456         if (sec > 0 && usec < 0) {
 457                 sec--;
 458                 usec += 1000000;
 459         }
 460
 461         if (sec < 0 || (sec == 0 && usec < 0))
 462                 return 0;
 463
 464         sec *= 1000;
 465         usec /= 1000;
 466         return sec + usec;
 467 }
 468
 469 uint64_t time_since_now(const struct timespec *s)
 470 {
 471         return mtime_since_now(s) / 1000;
 472 }
 473
 474 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 475     defined(CONFIG_SFAA)
 476
 477 #define CLOCK_ENTRIES_DEBUG     100000
 478 #define CLOCK_ENTRIES_TEST      10000
 479
 480 struct clock_entry {
 481         uint32_t seq;
 482         uint32_t cpu;
 483         uint64_t tsc;
 484 };
 485
 486 struct clock_thread {
 487         pthread_t thread;
 488         int cpu;
 489         int debug;
 490         pthread_mutex_t lock;
 491         pthread_mutex_t started;
 492         unsigned long nr_entries;
 493         uint32_t *seq;
 494         struct clock_entry *entries;
 495 };
 496
 497 static inline uint32_t atomic32_inc_return(uint32_t *seq)
 498 {
 499         return 1 + __sync_fetch_and_add(seq, 1);
 500 }
 501
 502 static void *clock_thread_fn(void *data)
 503 {
 504         struct clock_thread *t = data;
 505         struct clock_entry *c;
 506         os_cpu_mask_t cpu_mask;
 507         uint32_t last_seq;
 508         unsigned long long first;
 509         int i;
 510
 511         if (fio_cpuset_init(&cpu_mask)) {
 512                 int __err = errno;
 513
 514                 log_err("clock cpuset init failed: %s\n", strerror(__err));
 515                 goto err_out;
 516         }
 517
 518         fio_cpu_set(&cpu_mask, t->cpu);
 519
 520         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 521                 int __err = errno;
 522
 523                 log_err("clock setaffinity failed: %s\n", strerror(__err));
 524                 goto err;
 525         }
 526
 527         pthread_mutex_lock(&t->lock);
 528         pthread_mutex_unlock(&t->started);
 529
 530         first = get_cpu_clock();
 531         last_seq = 0;
 532         c = &t->entries[0];
 533         for (i = 0; i < t->nr_entries; i++, c++) {
 534                 uint32_t seq;
 535                 uint64_t tsc;
 536
 537                 c->cpu = t->cpu;
 538                 do {
 539                         seq = atomic32_inc_return(t->seq);
 540                         if (seq < last_seq)
 541                                 break;
 542                         tsc = get_cpu_clock();
 543                 } while (seq != *t->seq);
 544
 545                 c->seq = seq;
 546                 c->tsc = tsc;
 547         }
 548
 549         if (t->debug) {
 550                 unsigned long long clocks;
 551
 552                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
 553                 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
 554                                                         clocks, first);
 555         }
 556
 557         /*
 558          * The most common platform clock breakage is returning zero
 559          * indefinitely. Check for that and return failure.
 560          */
 561         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 562                 goto err;
 563
 564         fio_cpuset_exit(&cpu_mask);
 565         return NULL;
 566 err:
 567         fio_cpuset_exit(&cpu_mask);
 568 err_out:
 569         return (void *) 1;
 570 }
 571
 572 static int clock_cmp(const void *p1, const void *p2)
 573 {
 574         const struct clock_entry *c1 = p1;
 575         const struct clock_entry *c2 = p2;
 576
 577         if (c1->seq == c2->seq)
 578                 log_err("cs: bug in atomic sequence!\n");
 579
 580         return c1->seq - c2->seq;
 581 }
 582
 583 int fio_monotonic_clocktest(int debug)
 584 {
 585         struct clock_thread *cthreads;
 586         unsigned int nr_cpus = cpus_online();
 587         struct clock_entry *entries;
 588         unsigned long nr_entries, tentries, failed = 0;
 589         struct clock_entry *prev, *this;
 590         uint32_t seq = 0;
 591         unsigned int i;
 592
 593         if (debug) {
 594                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 595
 596 #ifdef FIO_INC_DEBUG
 597                 fio_debug |= 1U << FD_TIME;
 598 #endif
 599                 nr_entries = CLOCK_ENTRIES_DEBUG;
 600         } else
 601                 nr_entries = CLOCK_ENTRIES_TEST;
 602
 603         calibrate_cpu_clock();
 604
 605         if (debug) {
 606 #ifdef FIO_INC_DEBUG
 607                 fio_debug &= ~(1U << FD_TIME);
 608 #endif
 609         }
 610
 611         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 612         tentries = nr_entries * nr_cpus;
 613         entries = malloc(tentries * sizeof(struct clock_entry));
 614
 615         if (debug)
 616                 log_info("cs: Testing %u CPUs\n", nr_cpus);
 617
 618         for (i = 0; i < nr_cpus; i++) {
 619                 struct clock_thread *t = &cthreads[i];
 620
 621                 t->cpu = i;
 622                 t->debug = debug;
 623                 t->seq = &seq;
 624                 t->nr_entries = nr_entries;
 625                 t->entries = &entries[i * nr_entries];
 626                 pthread_mutex_init(&t->lock, NULL);
 627                 pthread_mutex_init(&t->started, NULL);
 628                 pthread_mutex_lock(&t->lock);
 629                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 630                         failed++;
 631                         nr_cpus = i;
 632                         break;
 633                 }
 634         }
 635
 636         for (i = 0; i < nr_cpus; i++) {
 637                 struct clock_thread *t = &cthreads[i];
 638
 639                 pthread_mutex_lock(&t->started);
 640         }
 641
 642         for (i = 0; i < nr_cpus; i++) {
 643                 struct clock_thread *t = &cthreads[i];
 644
 645                 pthread_mutex_unlock(&t->lock);
 646         }
 647
 648         for (i = 0; i < nr_cpus; i++) {
 649                 struct clock_thread *t = &cthreads[i];
 650                 void *ret;
 651
 652                 pthread_join(t->thread, &ret);
 653                 if (ret)
 654                         failed++;
 655         }
 656         free(cthreads);
 657
 658         if (failed) {
 659                 if (debug)
 660                         log_err("Clocksource test: %lu threads failed\n", failed);
 661                 goto err;
 662         }
 663
 664         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 665
 666         /* silence silly gcc */
 667         prev = NULL;
 668         for (failed = i = 0; i < tentries; i++) {
 669                 this = &entries[i];
 670
 671                 if (!i) {
 672                         prev = this;
 673                         continue;
 674                 }
 675
 676                 if (prev->tsc > this->tsc) {
 677                         uint64_t diff = prev->tsc - this->tsc;
 678
 679                         if (!debug) {
 680                                 failed++;
 681                                 break;
 682                         }
 683
 684                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 685                                                 (unsigned long long) diff);
 686                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 687                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 688                         failed++;
 689                 }
 690
 691                 prev = this;
 692         }
 693
 694         if (debug) {
 695                 if (failed)
 696                         log_info("cs: Failed: %lu\n", failed);
 697                 else
 698                         log_info("cs: Pass!\n");
 699         }
 700 err:
 701         free(entries);
 702         return !!failed;
 703 }
 704
 705 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 706
 707 int fio_monotonic_clocktest(int debug)
 708 {
 709         if (debug)
 710                 log_info("cs: current platform does not support CPU clocks\n");
 711         return 1;
 712 }
 713
 714 #endif