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_usec;
  20 static uint64_t max_cycles_for_mult;
  21 #endif
  22 #ifdef ARCH_CPU_CLOCK_WRAPS
  23 static unsigned long long cycles_start, cycles_wrap;
  24 #endif
  25 #endif
  26 int tsc_reliable = 0;
  27
  28 struct tv_valid {
  29         uint64_t last_cycles;
  30         int last_tv_valid;
  31         int warned;
  32 };
  33 #ifdef ARCH_HAVE_CPU_CLOCK
  34 #ifdef CONFIG_TLS_THREAD
  35 static __thread struct tv_valid static_tv_valid;
  36 #else
  37 static pthread_key_t tv_tls_key;
  38 #endif
  39 #endif
  40
  41 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  42 int fio_clock_source_set = 0;
  43 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  44
  45 #ifdef FIO_DEBUG_TIME
  46
  47 #define HASH_BITS       8
  48 #define HASH_SIZE       (1 << HASH_BITS)
  49
  50 static struct flist_head hash[HASH_SIZE];
  51 static int gtod_inited;
  52
  53 struct gtod_log {
  54         struct flist_head list;
  55         void *caller;
  56         unsigned long calls;
  57 };
  58
  59 static struct gtod_log *find_hash(void *caller)
  60 {
  61         unsigned long h = hash_ptr(caller, HASH_BITS);
  62         struct flist_head *entry;
  63
  64         flist_for_each(entry, &hash[h]) {
  65                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  66                                                                         list);
  67
  68                 if (log->caller == caller)
  69                         return log;
  70         }
  71
  72         return NULL;
  73 }
  74
  75 static void inc_caller(void *caller)
  76 {
  77         struct gtod_log *log = find_hash(caller);
  78
  79         if (!log) {
  80                 unsigned long h;
  81
  82                 log = malloc(sizeof(*log));
  83                 INIT_FLIST_HEAD(&log->list);
  84                 log->caller = caller;
  85                 log->calls = 0;
  86
  87                 h = hash_ptr(caller, HASH_BITS);
  88                 flist_add_tail(&log->list, &hash[h]);
  89         }
  90
  91         log->calls++;
  92 }
  93
  94 static void gtod_log_caller(void *caller)
  95 {
  96         if (gtod_inited)
  97                 inc_caller(caller);
  98 }
  99
 100 static void fio_exit fio_dump_gtod(void)
 101 {
 102         unsigned long total_calls = 0;
 103         int i;
 104
 105         for (i = 0; i < HASH_SIZE; i++) {
 106                 struct flist_head *entry;
 107                 struct gtod_log *log;
 108
 109                 flist_for_each(entry, &hash[i]) {
 110                         log = flist_entry(entry, struct gtod_log, list);
 111
 112                         printf("function %p, calls %lu\n", log->caller,
 113                                                                 log->calls);
 114                         total_calls += log->calls;
 115                 }
 116         }
 117
 118         printf("Total %lu gettimeofday\n", total_calls);
 119 }
 120
 121 static void fio_init gtod_init(void)
 122 {
 123         int i;
 124
 125         for (i = 0; i < HASH_SIZE; i++)
 126                 INIT_FLIST_HEAD(&hash[i]);
 127
 128         gtod_inited = 1;
 129 }
 130
 131 #endif /* FIO_DEBUG_TIME */
 132
 133 #ifdef CONFIG_CLOCK_GETTIME
 134 static int fill_clock_gettime(struct timespec *ts)
 135 {
 136 #ifdef CONFIG_CLOCK_MONOTONIC
 137         return clock_gettime(CLOCK_MONOTONIC, ts);
 138 #else
 139         return clock_gettime(CLOCK_REALTIME, ts);
 140 #endif
 141 }
 142 #endif
 143
 144 static void __fio_gettime(struct timeval *tp)
 145 {
 146         switch (fio_clock_source) {
 147 #ifdef CONFIG_GETTIMEOFDAY
 148         case CS_GTOD:
 149                 gettimeofday(tp, NULL);
 150                 break;
 151 #endif
 152 #ifdef CONFIG_CLOCK_GETTIME
 153         case CS_CGETTIME: {
 154                 struct timespec ts;
 155
 156                 if (fill_clock_gettime(&ts) < 0) {
 157                         log_err("fio: clock_gettime fails\n");
 158                         assert(0);
 159                 }
 160
 161                 tp->tv_sec = ts.tv_sec;
 162                 tp->tv_usec = ts.tv_nsec / 1000;
 163                 break;
 164                 }
 165 #endif
 166 #ifdef ARCH_HAVE_CPU_CLOCK
 167         case CS_CPUCLOCK: {
 168                 uint64_t usecs, t;
 169                 struct tv_valid *tv;
 170
 171 #ifdef CONFIG_TLS_THREAD
 172                 tv = &static_tv_valid;
 173 #else
 174                 tv = pthread_getspecific(tv_tls_key);
 175 #endif
 176
 177                 t = get_cpu_clock();
 178 #ifdef ARCH_CPU_CLOCK_WRAPS
 179                 if (t < cycles_start && !cycles_wrap)
 180                         cycles_wrap = 1;
 181                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
 182                         log_err("fio: double CPU clock wrap\n");
 183                         tv->warned = 1;
 184                 }
 185
 186                 t -= cycles_start;
 187 #endif
 188                 tv->last_cycles = t;
 189                 tv->last_tv_valid = 1;
 190 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 191                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 192 #else
 193                 if (t < max_cycles_for_mult)
 194                         usecs = (t * inv_cycles_per_usec) / 16777216UL;
 195                 else
 196                         usecs = t / cycles_per_usec;
 197 #endif
 198                 tp->tv_sec = usecs / 1000000;
 199                 tp->tv_usec = usecs % 1000000;
 200                 break;
 201                 }
 202 #endif
 203         default:
 204                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 205                 break;
 206         }
 207 }
 208
 209 #ifdef FIO_DEBUG_TIME
 210 void fio_gettime(struct timeval *tp, void *caller)
 211 #else
 212 void fio_gettime(struct timeval *tp, void fio_unused *caller)
 213 #endif
 214 {
 215 #ifdef FIO_DEBUG_TIME
 216         if (!caller)
 217                 caller = __builtin_return_address(0);
 218
 219         gtod_log_caller(caller);
 220 #endif
 221         if (fio_unlikely(fio_gettime_offload(tp)))
 222                 return;
 223
 224         __fio_gettime(tp);
 225 }
 226
 227 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 228 static unsigned long get_cycles_per_usec(void)
 229 {
 230         struct timeval s, e;
 231         uint64_t c_s, c_e;
 232         enum fio_cs old_cs = fio_clock_source;
 233
 234 #ifdef CONFIG_CLOCK_GETTIME
 235         fio_clock_source = CS_CGETTIME;
 236 #else
 237         fio_clock_source = CS_GTOD;
 238 #endif
 239         __fio_gettime(&s);
 240
 241         c_s = get_cpu_clock();
 242         do {
 243                 uint64_t elapsed;
 244
 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 + 127) >> 7;
 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;
 265
 266         cycles[0] = get_cycles_per_usec();
 267         S = delta = mean = 0.0;
 268         for (i = 0; i < NR_TIME_ITERS; i++) {
 269                 cycles[i] = get_cycles_per_usec();
 270                 delta = cycles[i] - mean;
 271                 if (delta) {
 272                         mean += delta / (i + 1.0);
 273                         S += delta * (cycles[i] - mean);
 274                 }
 275         }
 276
 277         /*
 278          * The most common platform clock breakage is returning zero
 279          * indefinitely. Check for that and return failure.
 280          */
 281         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 282                 return 1;
 283
 284         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 285
 286         minc = -1ULL;
 287         maxc = samples = avg = 0;
 288         for (i = 0; i < NR_TIME_ITERS; i++) {
 289                 double this = cycles[i];
 290
 291                 minc = min(cycles[i], minc);
 292                 maxc = max(cycles[i], maxc);
 293
 294                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 295                         continue;
 296                 samples++;
 297                 avg += this;
 298         }
 299
 300         S /= (double) NR_TIME_ITERS;
 301         mean /= 10.0;
 302
 303         for (i = 0; i < NR_TIME_ITERS; i++)
 304                 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
 305                                         (unsigned long long) cycles[i] / 10);
 306
 307         avg /= samples;
 308         avg = (avg + 5) / 10;
 309         minc /= 10;
 310         maxc /= 10;
 311         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 312         dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
 313                         (unsigned long long) minc,
 314                         (unsigned long long) maxc, mean, S);
 315
 316         cycles_per_usec = avg;
 317         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
 318         max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
 319         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
 320 #ifdef ARCH_CPU_CLOCK_WRAPS
 321         cycles_start = get_cpu_clock();
 322         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
 323 #endif
 324         return 0;
 325 }
 326 #else
 327 static int calibrate_cpu_clock(void)
 328 {
 329 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 330         return 0;
 331 #else
 332         return 1;
 333 #endif
 334 }
 335 #endif // ARCH_HAVE_CPU_CLOCK
 336
 337 #ifndef CONFIG_TLS_THREAD
 338 void fio_local_clock_init(int is_thread)
 339 {
 340         struct tv_valid *t;
 341
 342         t = calloc(1, sizeof(*t));
 343         if (pthread_setspecific(tv_tls_key, t)) {
 344                 log_err("fio: can't set TLS key\n");
 345                 assert(0);
 346         }
 347 }
 348
 349 static void kill_tv_tls_key(void *data)
 350 {
 351         free(data);
 352 }
 353 #else
 354 void fio_local_clock_init(int is_thread)
 355 {
 356 }
 357 #endif
 358
 359 void fio_clock_init(void)
 360 {
 361         if (fio_clock_source == fio_clock_source_inited)
 362                 return;
 363
 364 #ifndef CONFIG_TLS_THREAD
 365         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 366                 log_err("fio: can't create TLS key\n");
 367 #endif
 368
 369         fio_clock_source_inited = fio_clock_source;
 370
 371         if (calibrate_cpu_clock())
 372                 tsc_reliable = 0;
 373
 374         /*
 375          * If the arch sets tsc_reliable != 0, then it must be good enough
 376          * to use as THE clock source. For x86 CPUs, this means the TSC
 377          * runs at a constant rate and is synced across CPU cores.
 378          */
 379         if (tsc_reliable) {
 380                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
 381                         fio_clock_source = CS_CPUCLOCK;
 382         } else if (fio_clock_source == CS_CPUCLOCK)
 383                 log_info("fio: clocksource=cpu may not be reliable\n");
 384 }
 385
 386 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 387 {
 388         long sec, usec;
 389         uint64_t ret;
 390
 391         sec = e->tv_sec - s->tv_sec;
 392         usec = e->tv_usec - s->tv_usec;
 393         if (sec > 0 && usec < 0) {
 394                 sec--;
 395                 usec += 1000000;
 396         }
 397
 398         /*
 399          * time warp bug on some kernels?
 400          */
 401         if (sec < 0 || (sec == 0 && usec < 0))
 402                 return 0;
 403
 404         ret = sec * 1000000ULL + usec;
 405
 406         return ret;
 407 }
 408
 409 uint64_t utime_since_now(const struct timeval *s)
 410 {
 411         struct timeval t;
 412
 413         fio_gettime(&t, NULL);
 414         return utime_since(s, &t);
 415 }
 416
 417 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
 418 {
 419         long sec, usec, ret;
 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 *= 1000UL;
 432         usec /= 1000UL;
 433         ret = sec + usec;
 434
 435         return ret;
 436 }
 437
 438 uint64_t mtime_since_now(const struct timeval *s)
 439 {
 440         struct timeval t;
 441         void *p = __builtin_return_address(0);
 442
 443         fio_gettime(&t, p);
 444         return mtime_since(s, &t);
 445 }
 446
 447 uint64_t time_since_now(const struct timeval *s)
 448 {
 449         return mtime_since_now(s) / 1000;
 450 }
 451
 452 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 453     defined(CONFIG_SFAA)
 454
 455 #define CLOCK_ENTRIES_DEBUG     100000
 456 #define CLOCK_ENTRIES_TEST      10000
 457
 458 struct clock_entry {
 459         uint32_t seq;
 460         uint32_t cpu;
 461         uint64_t tsc;
 462 };
 463
 464 struct clock_thread {
 465         pthread_t thread;
 466         int cpu;
 467         int debug;
 468         pthread_mutex_t lock;
 469         pthread_mutex_t started;
 470         unsigned long nr_entries;
 471         uint32_t *seq;
 472         struct clock_entry *entries;
 473 };
 474
 475 static inline uint32_t atomic32_inc_return(uint32_t *seq)
 476 {
 477         return 1 + __sync_fetch_and_add(seq, 1);
 478 }
 479
 480 static void *clock_thread_fn(void *data)
 481 {
 482         struct clock_thread *t = data;
 483         struct clock_entry *c;
 484         os_cpu_mask_t cpu_mask;
 485         uint32_t last_seq;
 486         int i;
 487
 488         if (fio_cpuset_init(&cpu_mask)) {
 489                 int __err = errno;
 490
 491                 log_err("clock cpuset init failed: %s\n", strerror(__err));
 492                 goto err_out;
 493         }
 494
 495         fio_cpu_set(&cpu_mask, t->cpu);
 496
 497         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 498                 int __err = errno;
 499
 500                 log_err("clock setaffinity failed: %s\n", strerror(__err));
 501                 goto err;
 502         }
 503
 504         pthread_mutex_lock(&t->lock);
 505         pthread_mutex_unlock(&t->started);
 506
 507         last_seq = 0;
 508         c = &t->entries[0];
 509         for (i = 0; i < t->nr_entries; i++, c++) {
 510                 uint32_t seq;
 511                 uint64_t tsc;
 512
 513                 c->cpu = t->cpu;
 514                 do {
 515                         seq = atomic32_inc_return(t->seq);
 516                         if (seq < last_seq)
 517                                 break;
 518                         tsc = get_cpu_clock();
 519                 } while (seq != *t->seq);
 520
 521                 c->seq = seq;
 522                 c->tsc = tsc;
 523         }
 524
 525         if (t->debug) {
 526                 unsigned long long clocks;
 527
 528                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
 529                 log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu, clocks);
 530         }
 531
 532         /*
 533          * The most common platform clock breakage is returning zero
 534          * indefinitely. Check for that and return failure.
 535          */
 536         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 537                 goto err;
 538
 539         fio_cpuset_exit(&cpu_mask);
 540         return NULL;
 541 err:
 542         fio_cpuset_exit(&cpu_mask);
 543 err_out:
 544         return (void *) 1;
 545 }
 546
 547 static int clock_cmp(const void *p1, const void *p2)
 548 {
 549         const struct clock_entry *c1 = p1;
 550         const struct clock_entry *c2 = p2;
 551
 552         if (c1->seq == c2->seq)
 553                 log_err("cs: bug in atomic sequence!\n");
 554
 555         return c1->seq - c2->seq;
 556 }
 557
 558 int fio_monotonic_clocktest(int debug)
 559 {
 560         struct clock_thread *cthreads;
 561         unsigned int nr_cpus = cpus_online();
 562         struct clock_entry *entries;
 563         unsigned long nr_entries, tentries, failed = 0;
 564         struct clock_entry *prev, *this;
 565         uint32_t seq = 0;
 566         unsigned int i;
 567
 568         if (debug) {
 569                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 570
 571 #ifdef FIO_INC_DEBUG
 572                 fio_debug |= 1U << FD_TIME;
 573 #endif
 574                 nr_entries = CLOCK_ENTRIES_DEBUG;
 575         } else
 576                 nr_entries = CLOCK_ENTRIES_TEST;
 577
 578         calibrate_cpu_clock();
 579
 580         if (debug) {
 581 #ifdef FIO_INC_DEBUG
 582                 fio_debug &= ~(1U << FD_TIME);
 583 #endif
 584         }
 585
 586         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 587         tentries = nr_entries * nr_cpus;
 588         entries = malloc(tentries * sizeof(struct clock_entry));
 589
 590         if (debug)
 591                 log_info("cs: Testing %u CPUs\n", nr_cpus);
 592
 593         for (i = 0; i < nr_cpus; i++) {
 594                 struct clock_thread *t = &cthreads[i];
 595
 596                 t->cpu = i;
 597                 t->debug = debug;
 598                 t->seq = &seq;
 599                 t->nr_entries = nr_entries;
 600                 t->entries = &entries[i * nr_entries];
 601                 pthread_mutex_init(&t->lock, NULL);
 602                 pthread_mutex_init(&t->started, NULL);
 603                 pthread_mutex_lock(&t->lock);
 604                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 605                         failed++;
 606                         nr_cpus = i;
 607                         break;
 608                 }
 609         }
 610
 611         for (i = 0; i < nr_cpus; i++) {
 612                 struct clock_thread *t = &cthreads[i];
 613
 614                 pthread_mutex_lock(&t->started);
 615         }
 616
 617         for (i = 0; i < nr_cpus; i++) {
 618                 struct clock_thread *t = &cthreads[i];
 619
 620                 pthread_mutex_unlock(&t->lock);
 621         }
 622
 623         for (i = 0; i < nr_cpus; i++) {
 624                 struct clock_thread *t = &cthreads[i];
 625                 void *ret;
 626
 627                 pthread_join(t->thread, &ret);
 628                 if (ret)
 629                         failed++;
 630         }
 631         free(cthreads);
 632
 633         if (failed) {
 634                 if (debug)
 635                         log_err("Clocksource test: %lu threads failed\n", failed);
 636                 goto err;
 637         }
 638
 639         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 640
 641         for (failed = i = 0; i < tentries; i++) {
 642                 this = &entries[i];
 643
 644                 if (!i) {
 645                         prev = this;
 646                         continue;
 647                 }
 648
 649                 if (prev->tsc > this->tsc) {
 650                         uint64_t diff = prev->tsc - this->tsc;
 651
 652                         if (!debug) {
 653                                 failed++;
 654                                 break;
 655                         }
 656
 657                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 658                                                 (unsigned long long) diff);
 659                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 660                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 661                         failed++;
 662                 }
 663
 664                 prev = this;
 665         }
 666
 667         if (debug) {
 668                 if (failed)
 669                         log_info("cs: Failed: %lu\n", failed);
 670                 else
 671                         log_info("cs: Pass!\n");
 672         }
 673 err:
 674         free(entries);
 675         return !!failed;
 676 }
 677
 678 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 679
 680 int fio_monotonic_clocktest(int debug)
 681 {
 682         if (debug)
 683                 log_info("cs: current platform does not support CPU clocks\n");
 684         return 1;
 685 }
 686
 687 #endif