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 #if defined(CONFIG_CLOCK_MONOTONIC_RAW)
 137         return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
 138 #elif defined(CONFIG_CLOCK_MONOTONIC)
 139         return clock_gettime(CLOCK_MONOTONIC, ts);
 140 #else
 141         return clock_gettime(CLOCK_REALTIME, ts);
 142 #endif
 143 }
 144 #endif
 145
 146 static void __fio_gettime(struct timeval *tp)
 147 {
 148         switch (fio_clock_source) {
 149 #ifdef CONFIG_GETTIMEOFDAY
 150         case CS_GTOD:
 151                 gettimeofday(tp, NULL);
 152                 break;
 153 #endif
 154 #ifdef CONFIG_CLOCK_GETTIME
 155         case CS_CGETTIME: {
 156                 struct timespec ts;
 157
 158                 if (fill_clock_gettime(&ts) < 0) {
 159                         log_err("fio: clock_gettime fails\n");
 160                         assert(0);
 161                 }
 162
 163                 tp->tv_sec = ts.tv_sec;
 164                 tp->tv_usec = ts.tv_nsec / 1000;
 165                 break;
 166                 }
 167 #endif
 168 #ifdef ARCH_HAVE_CPU_CLOCK
 169         case CS_CPUCLOCK: {
 170                 uint64_t usecs, t;
 171                 struct tv_valid *tv;
 172
 173 #ifdef CONFIG_TLS_THREAD
 174                 tv = &static_tv_valid;
 175 #else
 176                 tv = pthread_getspecific(tv_tls_key);
 177 #endif
 178
 179                 t = get_cpu_clock();
 180 #ifdef ARCH_CPU_CLOCK_WRAPS
 181                 if (t < cycles_start && !cycles_wrap)
 182                         cycles_wrap = 1;
 183                 else if (cycles_wrap && t >= cycles_start && !tv->warned) {
 184                         log_err("fio: double CPU clock wrap\n");
 185                         tv->warned = 1;
 186                 }
 187
 188                 t -= cycles_start;
 189 #endif
 190                 tv->last_cycles = t;
 191                 tv->last_tv_valid = 1;
 192 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 193                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 194 #else
 195                 if (t < max_cycles_for_mult)
 196                         usecs = (t * inv_cycles_per_usec) / 16777216UL;
 197                 else
 198                         usecs = t / cycles_per_usec;
 199 #endif
 200                 tp->tv_sec = usecs / 1000000;
 201                 tp->tv_usec = usecs % 1000000;
 202                 break;
 203                 }
 204 #endif
 205         default:
 206                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 207                 break;
 208         }
 209 }
 210
 211 #ifdef FIO_DEBUG_TIME
 212 void fio_gettime(struct timeval *tp, void *caller)
 213 #else
 214 void fio_gettime(struct timeval *tp, void fio_unused *caller)
 215 #endif
 216 {
 217 #ifdef FIO_DEBUG_TIME
 218         if (!caller)
 219                 caller = __builtin_return_address(0);
 220
 221         gtod_log_caller(caller);
 222 #endif
 223         if (fio_unlikely(fio_gettime_offload(tp)))
 224                 return;
 225
 226         __fio_gettime(tp);
 227 }
 228
 229 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 230 static unsigned long get_cycles_per_usec(void)
 231 {
 232         struct timeval s, e;
 233         uint64_t c_s, c_e;
 234         enum fio_cs old_cs = fio_clock_source;
 235         uint64_t elapsed;
 236
 237 #ifdef CONFIG_CLOCK_GETTIME
 238         fio_clock_source = CS_CGETTIME;
 239 #else
 240         fio_clock_source = CS_GTOD;
 241 #endif
 242         __fio_gettime(&s);
 243
 244         c_s = get_cpu_clock();
 245         do {
 246                 __fio_gettime(&e);
 247
 248                 elapsed = utime_since(&s, &e);
 249                 if (elapsed >= 1280) {
 250                         c_e = get_cpu_clock();
 251                         break;
 252                 }
 253         } while (1);
 254
 255         fio_clock_source = old_cs;
 256         return (c_e - c_s) / elapsed;
 257 }
 258
 259 #define NR_TIME_ITERS   50
 260
 261 static int calibrate_cpu_clock(void)
 262 {
 263         double delta, mean, S;
 264         uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
 265         int i, samples;
 266
 267         cycles[0] = get_cycles_per_usec();
 268         S = delta = mean = 0.0;
 269         for (i = 0; i < NR_TIME_ITERS; i++) {
 270                 cycles[i] = get_cycles_per_usec();
 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         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 308         dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
 309                         (unsigned long long) minc,
 310                         (unsigned long long) maxc, mean, S);
 311
 312         cycles_per_usec = avg;
 313         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
 314         max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
 315         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
 316 #ifdef ARCH_CPU_CLOCK_WRAPS
 317         cycles_start = get_cpu_clock();
 318         dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
 319 #endif
 320         return 0;
 321 }
 322 #else
 323 static int calibrate_cpu_clock(void)
 324 {
 325 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 326         return 0;
 327 #else
 328         return 1;
 329 #endif
 330 }
 331 #endif // ARCH_HAVE_CPU_CLOCK
 332
 333 #ifndef CONFIG_TLS_THREAD
 334 void fio_local_clock_init(int is_thread)
 335 {
 336         struct tv_valid *t;
 337
 338         t = calloc(1, sizeof(*t));
 339         if (pthread_setspecific(tv_tls_key, t)) {
 340                 log_err("fio: can't set TLS key\n");
 341                 assert(0);
 342         }
 343 }
 344
 345 static void kill_tv_tls_key(void *data)
 346 {
 347         free(data);
 348 }
 349 #else
 350 void fio_local_clock_init(int is_thread)
 351 {
 352 }
 353 #endif
 354
 355 void fio_clock_init(void)
 356 {
 357         if (fio_clock_source == fio_clock_source_inited)
 358                 return;
 359
 360 #ifndef CONFIG_TLS_THREAD
 361         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 362                 log_err("fio: can't create TLS key\n");
 363 #endif
 364
 365         fio_clock_source_inited = fio_clock_source;
 366
 367         if (calibrate_cpu_clock())
 368                 tsc_reliable = 0;
 369
 370         /*
 371          * If the arch sets tsc_reliable != 0, then it must be good enough
 372          * to use as THE clock source. For x86 CPUs, this means the TSC
 373          * runs at a constant rate and is synced across CPU cores.
 374          */
 375         if (tsc_reliable) {
 376                 if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
 377                         fio_clock_source = CS_CPUCLOCK;
 378         } else if (fio_clock_source == CS_CPUCLOCK)
 379                 log_info("fio: clocksource=cpu may not be reliable\n");
 380 }
 381
 382 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 383 {
 384         int64_t sec, usec;
 385
 386         sec = e->tv_sec - s->tv_sec;
 387         usec = e->tv_usec - s->tv_usec;
 388         if (sec > 0 && usec < 0) {
 389                 sec--;
 390                 usec += 1000000;
 391         }
 392
 393         /*
 394          * time warp bug on some kernels?
 395          */
 396         if (sec < 0 || (sec == 0 && usec < 0))
 397                 return 0;
 398
 399         return usec + (sec * 1000000);
 400 }
 401
 402 uint64_t utime_since_now(const struct timeval *s)
 403 {
 404         struct timeval t;
 405
 406         fio_gettime(&t, NULL);
 407         return utime_since(s, &t);
 408 }
 409
 410 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
 411 {
 412         long sec, usec;
 413
 414         sec = e->tv_sec - s->tv_sec;
 415         usec = e->tv_usec - s->tv_usec;
 416         if (sec > 0 && usec < 0) {
 417                 sec--;
 418                 usec += 1000000;
 419         }
 420
 421         if (sec < 0 || (sec == 0 && usec < 0))
 422                 return 0;
 423
 424         sec *= 1000;
 425         usec /= 1000;
 426         return sec + usec;
 427 }
 428
 429 uint64_t mtime_since_now(const struct timeval *s)
 430 {
 431         struct timeval t;
 432         void *p = __builtin_return_address(0);
 433
 434         fio_gettime(&t, p);
 435         return mtime_since(s, &t);
 436 }
 437
 438 uint64_t time_since_now(const struct timeval *s)
 439 {
 440         return mtime_since_now(s) / 1000;
 441 }
 442
 443 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 444     defined(CONFIG_SFAA)
 445
 446 #define CLOCK_ENTRIES_DEBUG     100000
 447 #define CLOCK_ENTRIES_TEST      10000
 448
 449 struct clock_entry {
 450         uint32_t seq;
 451         uint32_t cpu;
 452         uint64_t tsc;
 453 };
 454
 455 struct clock_thread {
 456         pthread_t thread;
 457         int cpu;
 458         int debug;
 459         pthread_mutex_t lock;
 460         pthread_mutex_t started;
 461         unsigned long nr_entries;
 462         uint32_t *seq;
 463         struct clock_entry *entries;
 464 };
 465
 466 static inline uint32_t atomic32_inc_return(uint32_t *seq)
 467 {
 468         return 1 + __sync_fetch_and_add(seq, 1);
 469 }
 470
 471 static void *clock_thread_fn(void *data)
 472 {
 473         struct clock_thread *t = data;
 474         struct clock_entry *c;
 475         os_cpu_mask_t cpu_mask;
 476         uint32_t last_seq;
 477         unsigned long long first;
 478         int i;
 479
 480         if (fio_cpuset_init(&cpu_mask)) {
 481                 int __err = errno;
 482
 483                 log_err("clock cpuset init failed: %s\n", strerror(__err));
 484                 goto err_out;
 485         }
 486
 487         fio_cpu_set(&cpu_mask, t->cpu);
 488
 489         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 490                 int __err = errno;
 491
 492                 log_err("clock setaffinity failed: %s\n", strerror(__err));
 493                 goto err;
 494         }
 495
 496         pthread_mutex_lock(&t->lock);
 497         pthread_mutex_unlock(&t->started);
 498
 499         first = get_cpu_clock();
 500         last_seq = 0;
 501         c = &t->entries[0];
 502         for (i = 0; i < t->nr_entries; i++, c++) {
 503                 uint32_t seq;
 504                 uint64_t tsc;
 505
 506                 c->cpu = t->cpu;
 507                 do {
 508                         seq = atomic32_inc_return(t->seq);
 509                         if (seq < last_seq)
 510                                 break;
 511                         tsc = get_cpu_clock();
 512                 } while (seq != *t->seq);
 513
 514                 c->seq = seq;
 515                 c->tsc = tsc;
 516         }
 517
 518         if (t->debug) {
 519                 unsigned long long clocks;
 520
 521                 clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
 522                 log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu,
 523                                                         clocks, first);
 524         }
 525
 526         /*
 527          * The most common platform clock breakage is returning zero
 528          * indefinitely. Check for that and return failure.
 529          */
 530         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 531                 goto err;
 532
 533         fio_cpuset_exit(&cpu_mask);
 534         return NULL;
 535 err:
 536         fio_cpuset_exit(&cpu_mask);
 537 err_out:
 538         return (void *) 1;
 539 }
 540
 541 static int clock_cmp(const void *p1, const void *p2)
 542 {
 543         const struct clock_entry *c1 = p1;
 544         const struct clock_entry *c2 = p2;
 545
 546         if (c1->seq == c2->seq)
 547                 log_err("cs: bug in atomic sequence!\n");
 548
 549         return c1->seq - c2->seq;
 550 }
 551
 552 int fio_monotonic_clocktest(int debug)
 553 {
 554         struct clock_thread *cthreads;
 555         unsigned int nr_cpus = cpus_online();
 556         struct clock_entry *entries;
 557         unsigned long nr_entries, tentries, failed = 0;
 558         struct clock_entry *prev, *this;
 559         uint32_t seq = 0;
 560         unsigned int i;
 561
 562         if (debug) {
 563                 log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 564
 565 #ifdef FIO_INC_DEBUG
 566                 fio_debug |= 1U << FD_TIME;
 567 #endif
 568                 nr_entries = CLOCK_ENTRIES_DEBUG;
 569         } else
 570                 nr_entries = CLOCK_ENTRIES_TEST;
 571
 572         calibrate_cpu_clock();
 573
 574         if (debug) {
 575 #ifdef FIO_INC_DEBUG
 576                 fio_debug &= ~(1U << FD_TIME);
 577 #endif
 578         }
 579
 580         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 581         tentries = nr_entries * nr_cpus;
 582         entries = malloc(tentries * sizeof(struct clock_entry));
 583
 584         if (debug)
 585                 log_info("cs: Testing %u CPUs\n", nr_cpus);
 586
 587         for (i = 0; i < nr_cpus; i++) {
 588                 struct clock_thread *t = &cthreads[i];
 589
 590                 t->cpu = i;
 591                 t->debug = debug;
 592                 t->seq = &seq;
 593                 t->nr_entries = nr_entries;
 594                 t->entries = &entries[i * nr_entries];
 595                 pthread_mutex_init(&t->lock, NULL);
 596                 pthread_mutex_init(&t->started, NULL);
 597                 pthread_mutex_lock(&t->lock);
 598                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 599                         failed++;
 600                         nr_cpus = i;
 601                         break;
 602                 }
 603         }
 604
 605         for (i = 0; i < nr_cpus; i++) {
 606                 struct clock_thread *t = &cthreads[i];
 607
 608                 pthread_mutex_lock(&t->started);
 609         }
 610
 611         for (i = 0; i < nr_cpus; i++) {
 612                 struct clock_thread *t = &cthreads[i];
 613
 614                 pthread_mutex_unlock(&t->lock);
 615         }
 616
 617         for (i = 0; i < nr_cpus; i++) {
 618                 struct clock_thread *t = &cthreads[i];
 619                 void *ret;
 620
 621                 pthread_join(t->thread, &ret);
 622                 if (ret)
 623                         failed++;
 624         }
 625         free(cthreads);
 626
 627         if (failed) {
 628                 if (debug)
 629                         log_err("Clocksource test: %lu threads failed\n", failed);
 630                 goto err;
 631         }
 632
 633         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 634
 635         /* silence silly gcc */
 636         prev = NULL;
 637         for (failed = i = 0; i < tentries; i++) {
 638                 this = &entries[i];
 639
 640                 if (!i) {
 641                         prev = this;
 642                         continue;
 643                 }
 644
 645                 if (prev->tsc > this->tsc) {
 646                         uint64_t diff = prev->tsc - this->tsc;
 647
 648                         if (!debug) {
 649                                 failed++;
 650                                 break;
 651                         }
 652
 653                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 654                                                 (unsigned long long) diff);
 655                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 656                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 657                         failed++;
 658                 }
 659
 660                 prev = this;
 661         }
 662
 663         if (debug) {
 664                 if (failed)
 665                         log_info("cs: Failed: %lu\n", failed);
 666                 else
 667                         log_info("cs: Pass!\n");
 668         }
 669 err:
 670         free(entries);
 671         return !!failed;
 672 }
 673
 674 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 675
 676 int fio_monotonic_clocktest(int debug)
 677 {
 678         if (debug)
 679                 log_info("cs: current platform does not support CPU clocks\n");
 680         return 1;
 681 }
 682
 683 #endif