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) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
  17 static unsigned long cycles_per_usec;
  18 static unsigned long inv_cycles_per_usec;
  19 static uint64_t max_cycles_for_mult;
  20 #endif
  21 int tsc_reliable = 0;
  22
  23 struct tv_valid {
  24         uint64_t last_cycles;
  25         int last_tv_valid;
  26         int warned;
  27 };
  28 #ifdef CONFIG_TLS_THREAD
  29 static __thread struct tv_valid static_tv_valid;
  30 #else
  31 static pthread_key_t tv_tls_key;
  32 #endif
  33
  34 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  35 int fio_clock_source_set = 0;
  36 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  37
  38 #ifdef FIO_DEBUG_TIME
  39
  40 #define HASH_BITS       8
  41 #define HASH_SIZE       (1 << HASH_BITS)
  42
  43 static struct flist_head hash[HASH_SIZE];
  44 static int gtod_inited;
  45
  46 struct gtod_log {
  47         struct flist_head list;
  48         void *caller;
  49         unsigned long calls;
  50 };
  51
  52 static struct gtod_log *find_hash(void *caller)
  53 {
  54         unsigned long h = hash_ptr(caller, HASH_BITS);
  55         struct flist_head *entry;
  56
  57         flist_for_each(entry, &hash[h]) {
  58                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  59                                                                         list);
  60
  61                 if (log->caller == caller)
  62                         return log;
  63         }
  64
  65         return NULL;
  66 }
  67
  68 static struct gtod_log *find_log(void *caller)
  69 {
  70         struct gtod_log *log = find_hash(caller);
  71
  72         if (!log) {
  73                 unsigned long h;
  74
  75                 log = malloc(sizeof(*log));
  76                 INIT_FLIST_HEAD(&log->list);
  77                 log->caller = caller;
  78                 log->calls = 0;
  79
  80                 h = hash_ptr(caller, HASH_BITS);
  81                 flist_add_tail(&log->list, &hash[h]);
  82         }
  83
  84         return log;
  85 }
  86
  87 static void gtod_log_caller(void *caller)
  88 {
  89         if (gtod_inited) {
  90                 struct gtod_log *log = find_log(caller);
  91
  92                 log->calls++;
  93         }
  94 }
  95
  96 static void fio_exit fio_dump_gtod(void)
  97 {
  98         unsigned long total_calls = 0;
  99         int i;
 100
 101         for (i = 0; i < HASH_SIZE; i++) {
 102                 struct flist_head *entry;
 103                 struct gtod_log *log;
 104
 105                 flist_for_each(entry, &hash[i]) {
 106                         log = flist_entry(entry, struct gtod_log, list);
 107
 108                         printf("function %p, calls %lu\n", log->caller,
 109                                                                 log->calls);
 110                         total_calls += log->calls;
 111                 }
 112         }
 113
 114         printf("Total %lu gettimeofday\n", total_calls);
 115 }
 116
 117 static void fio_init gtod_init(void)
 118 {
 119         int i;
 120
 121         for (i = 0; i < HASH_SIZE; i++)
 122                 INIT_FLIST_HEAD(&hash[i]);
 123
 124         gtod_inited = 1;
 125 }
 126
 127 #endif /* FIO_DEBUG_TIME */
 128
 129 #ifdef CONFIG_CLOCK_GETTIME
 130 static int fill_clock_gettime(struct timespec *ts)
 131 {
 132 #ifdef CONFIG_CLOCK_MONOTONIC
 133         return clock_gettime(CLOCK_MONOTONIC, ts);
 134 #else
 135         return clock_gettime(CLOCK_REALTIME, ts);
 136 #endif
 137 }
 138 #endif
 139
 140 static void __fio_gettime(struct timeval *tp)
 141 {
 142         struct tv_valid *tv;
 143
 144 #ifdef CONFIG_TLS_THREAD
 145         tv = &static_tv_valid;
 146 #else
 147         tv = pthread_getspecific(tv_tls_key);
 148 #endif
 149
 150         switch (fio_clock_source) {
 151 #ifdef CONFIG_GETTIMEOFDAY
 152         case CS_GTOD:
 153                 gettimeofday(tp, NULL);
 154                 break;
 155 #endif
 156 #ifdef CONFIG_CLOCK_GETTIME
 157         case CS_CGETTIME: {
 158                 struct timespec ts;
 159
 160                 if (fill_clock_gettime(&ts) < 0) {
 161                         log_err("fio: clock_gettime fails\n");
 162                         assert(0);
 163                 }
 164
 165                 tp->tv_sec = ts.tv_sec;
 166                 tp->tv_usec = ts.tv_nsec / 1000;
 167                 break;
 168                 }
 169 #endif
 170 #ifdef ARCH_HAVE_CPU_CLOCK
 171         case CS_CPUCLOCK: {
 172                 uint64_t usecs, t;
 173
 174                 t = get_cpu_clock();
 175                 if (t < tv->last_cycles && tv->last_tv_valid &&
 176                     !tv->warned) {
 177                         log_err("fio: CPU clock going back in time\n");
 178                         tv->warned = 1;
 179                 }
 180
 181                 tv->last_cycles = t;
 182                 tv->last_tv_valid = 1;
 183 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 184                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 185 #else
 186                 if (t < max_cycles_for_mult)
 187                         usecs = (t * inv_cycles_per_usec) / 16777216UL;
 188                 else
 189                         usecs = t / cycles_per_usec;
 190 #endif
 191                 tp->tv_sec = usecs / 1000000;
 192                 tp->tv_usec = usecs % 1000000;
 193                 break;
 194                 }
 195 #endif
 196         default:
 197                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 198                 break;
 199         }
 200 }
 201
 202 #ifdef FIO_DEBUG_TIME
 203 void fio_gettime(struct timeval *tp, void *caller)
 204 #else
 205 void fio_gettime(struct timeval *tp, void fio_unused *caller)
 206 #endif
 207 {
 208 #ifdef FIO_DEBUG_TIME
 209         if (!caller)
 210                 caller = __builtin_return_address(0);
 211
 212         gtod_log_caller(caller);
 213 #endif
 214         if (fio_unlikely(fio_gettime_offload(tp)))
 215                 return;
 216
 217         __fio_gettime(tp);
 218 }
 219
 220 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 221 static unsigned long get_cycles_per_usec(void)
 222 {
 223         struct timeval s, e;
 224         uint64_t c_s, c_e;
 225         enum fio_cs old_cs = fio_clock_source;
 226
 227 #ifdef CONFIG_CLOCK_GETTIME
 228         fio_clock_source = CS_CGETTIME;
 229 #else
 230         fio_clock_source = CS_GTOD;
 231 #endif
 232         __fio_gettime(&s);
 233
 234         c_s = get_cpu_clock();
 235         do {
 236                 uint64_t elapsed;
 237
 238                 __fio_gettime(&e);
 239
 240                 elapsed = utime_since(&s, &e);
 241                 if (elapsed >= 1280) {
 242                         c_e = get_cpu_clock();
 243                         break;
 244                 }
 245         } while (1);
 246
 247         fio_clock_source = old_cs;
 248         return (c_e - c_s + 127) >> 7;
 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 avg, cycles[NR_TIME_ITERS];
 257         int i, samples;
 258
 259         cycles[0] = get_cycles_per_usec();
 260         S = delta = mean = 0.0;
 261         for (i = 0; i < NR_TIME_ITERS; i++) {
 262                 cycles[i] = get_cycles_per_usec();
 263                 delta = cycles[i] - mean;
 264                 if (delta) {
 265                         mean += delta / (i + 1.0);
 266                         S += delta * (cycles[i] - mean);
 267                 }
 268         }
 269
 270         /*
 271          * The most common platform clock breakage is returning zero
 272          * indefinitely. Check for that and return failure.
 273          */
 274         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 275                 return 1;
 276
 277         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 278
 279         samples = avg = 0;
 280         for (i = 0; i < NR_TIME_ITERS; i++) {
 281                 double this = cycles[i];
 282
 283                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 284                         continue;
 285                 samples++;
 286                 avg += this;
 287         }
 288
 289         S /= (double) NR_TIME_ITERS;
 290         mean /= 10.0;
 291
 292         for (i = 0; i < NR_TIME_ITERS; i++)
 293                 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
 294                                         (unsigned long long) cycles[i] / 10);
 295
 296         avg /= samples;
 297         avg = (avg + 5) / 10;
 298         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 299         dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
 300
 301         cycles_per_usec = avg;
 302         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
 303         max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
 304         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
 305         return 0;
 306 }
 307 #else
 308 static int calibrate_cpu_clock(void)
 309 {
 310 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 311         return 0;
 312 #else
 313         return 1;
 314 #endif
 315 }
 316 #endif // ARCH_HAVE_CPU_CLOCK
 317
 318 #ifndef CONFIG_TLS_THREAD
 319 void fio_local_clock_init(int is_thread)
 320 {
 321         struct tv_valid *t;
 322
 323         t = calloc(1, sizeof(*t));
 324         if (pthread_setspecific(tv_tls_key, t)) {
 325                 log_err("fio: can't set TLS key\n");
 326                 assert(0);
 327         }
 328 }
 329
 330 static void kill_tv_tls_key(void *data)
 331 {
 332         free(data);
 333 }
 334 #else
 335 void fio_local_clock_init(int is_thread)
 336 {
 337 }
 338 #endif
 339
 340 void fio_clock_init(void)
 341 {
 342         if (fio_clock_source == fio_clock_source_inited)
 343                 return;
 344
 345 #ifndef CONFIG_TLS_THREAD
 346         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 347                 log_err("fio: can't create TLS key\n");
 348 #endif
 349
 350         fio_clock_source_inited = fio_clock_source;
 351
 352         if (calibrate_cpu_clock())
 353                 tsc_reliable = 0;
 354
 355         /*
 356          * If the arch sets tsc_reliable != 0, then it must be good enough
 357          * to use as THE clock source. For x86 CPUs, this means the TSC
 358          * runs at a constant rate and is synced across CPU cores.
 359          */
 360         if (tsc_reliable) {
 361                 if (!fio_clock_source_set)
 362                         fio_clock_source = CS_CPUCLOCK;
 363         } else if (fio_clock_source == CS_CPUCLOCK)
 364                 log_info("fio: clocksource=cpu may not be reliable\n");
 365 }
 366
 367 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 368 {
 369         long sec, usec;
 370         uint64_t ret;
 371
 372         sec = e->tv_sec - s->tv_sec;
 373         usec = e->tv_usec - s->tv_usec;
 374         if (sec > 0 && usec < 0) {
 375                 sec--;
 376                 usec += 1000000;
 377         }
 378
 379         /*
 380          * time warp bug on some kernels?
 381          */
 382         if (sec < 0 || (sec == 0 && usec < 0))
 383                 return 0;
 384
 385         ret = sec * 1000000ULL + usec;
 386
 387         return ret;
 388 }
 389
 390 uint64_t utime_since_now(const struct timeval *s)
 391 {
 392         struct timeval t;
 393
 394         fio_gettime(&t, NULL);
 395         return utime_since(s, &t);
 396 }
 397
 398 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
 399 {
 400         long sec, usec, ret;
 401
 402         sec = e->tv_sec - s->tv_sec;
 403         usec = e->tv_usec - s->tv_usec;
 404         if (sec > 0 && usec < 0) {
 405                 sec--;
 406                 usec += 1000000;
 407         }
 408
 409         if (sec < 0 || (sec == 0 && usec < 0))
 410                 return 0;
 411
 412         sec *= 1000UL;
 413         usec /= 1000UL;
 414         ret = sec + usec;
 415
 416         return ret;
 417 }
 418
 419 uint64_t mtime_since_now(const struct timeval *s)
 420 {
 421         struct timeval t;
 422         void *p = __builtin_return_address(0);
 423
 424         fio_gettime(&t, p);
 425         return mtime_since(s, &t);
 426 }
 427
 428 uint64_t time_since_now(const struct timeval *s)
 429 {
 430         return mtime_since_now(s) / 1000;
 431 }
 432
 433 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 434     defined(CONFIG_SFAA)
 435
 436 #define CLOCK_ENTRIES   100000
 437
 438 struct clock_entry {
 439         uint32_t seq;
 440         uint32_t cpu;
 441         uint64_t tsc;
 442 };
 443
 444 struct clock_thread {
 445         pthread_t thread;
 446         int cpu;
 447         pthread_mutex_t lock;
 448         pthread_mutex_t started;
 449         uint32_t *seq;
 450         struct clock_entry *entries;
 451 };
 452
 453 static inline uint32_t atomic32_inc_return(uint32_t *seq)
 454 {
 455         return 1 + __sync_fetch_and_add(seq, 1);
 456 }
 457
 458 static void *clock_thread_fn(void *data)
 459 {
 460         struct clock_thread *t = data;
 461         struct clock_entry *c;
 462         os_cpu_mask_t cpu_mask;
 463         uint32_t last_seq;
 464         int i;
 465
 466         memset(&cpu_mask, 0, sizeof(cpu_mask));
 467         fio_cpu_set(&cpu_mask, t->cpu);
 468
 469         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 470                 log_err("clock setaffinity failed\n");
 471                 return (void *) 1;
 472         }
 473
 474         pthread_mutex_lock(&t->lock);
 475         pthread_mutex_unlock(&t->started);
 476
 477         last_seq = 0;
 478         c = &t->entries[0];
 479         for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
 480                 uint32_t seq;
 481                 uint64_t tsc;
 482
 483                 c->cpu = t->cpu;
 484                 do {
 485                         seq = atomic32_inc_return(t->seq);
 486                         if (seq < last_seq)
 487                                 break;
 488                         tsc = get_cpu_clock();
 489                 } while (seq != *t->seq);
 490
 491                 c->seq = seq;
 492                 c->tsc = tsc;
 493         }
 494
 495         log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
 496                 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
 497
 498         /*
 499          * The most common platform clock breakage is returning zero
 500          * indefinitely. Check for that and return failure.
 501          */
 502         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 503                 return (void *) 1;
 504
 505         return NULL;
 506 }
 507
 508 static int clock_cmp(const void *p1, const void *p2)
 509 {
 510         const struct clock_entry *c1 = p1;
 511         const struct clock_entry *c2 = p2;
 512
 513         if (c1->seq == c2->seq)
 514                 log_err("cs: bug in atomic sequence!\n");
 515
 516         return c1->seq - c2->seq;
 517 }
 518
 519 int fio_monotonic_clocktest(void)
 520 {
 521         struct clock_thread *cthreads;
 522         unsigned int nr_cpus = cpus_online();
 523         struct clock_entry *entries;
 524         unsigned long tentries, failed = 0;
 525         struct clock_entry *prev, *this;
 526         uint32_t seq = 0;
 527         unsigned int i;
 528
 529         log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 530
 531 #ifdef FIO_INC_DEBUG
 532         fio_debug |= 1U << FD_TIME;
 533 #endif
 534         calibrate_cpu_clock();
 535 #ifdef FIO_INC_DEBUG
 536         fio_debug &= ~(1U << FD_TIME);
 537 #endif
 538
 539         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 540         tentries = CLOCK_ENTRIES * nr_cpus;
 541         entries = malloc(tentries * sizeof(struct clock_entry));
 542
 543         log_info("cs: Testing %u CPUs\n", nr_cpus);
 544
 545         for (i = 0; i < nr_cpus; i++) {
 546                 struct clock_thread *t = &cthreads[i];
 547
 548                 t->cpu = i;
 549                 t->seq = &seq;
 550                 t->entries = &entries[i * CLOCK_ENTRIES];
 551                 pthread_mutex_init(&t->lock, NULL);
 552                 pthread_mutex_init(&t->started, NULL);
 553                 pthread_mutex_lock(&t->lock);
 554                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 555                         failed++;
 556                         nr_cpus = i;
 557                         break;
 558                 }
 559         }
 560
 561         for (i = 0; i < nr_cpus; i++) {
 562                 struct clock_thread *t = &cthreads[i];
 563
 564                 pthread_mutex_lock(&t->started);
 565         }
 566
 567         for (i = 0; i < nr_cpus; i++) {
 568                 struct clock_thread *t = &cthreads[i];
 569
 570                 pthread_mutex_unlock(&t->lock);
 571         }
 572
 573         for (i = 0; i < nr_cpus; i++) {
 574                 struct clock_thread *t = &cthreads[i];
 575                 void *ret;
 576
 577                 pthread_join(t->thread, &ret);
 578                 if (ret)
 579                         failed++;
 580         }
 581         free(cthreads);
 582
 583         if (failed) {
 584                 log_err("Clocksource test: %lu threads failed\n", failed);
 585                 goto err;
 586         }
 587
 588         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 589
 590         for (failed = i = 0; i < tentries; i++) {
 591                 this = &entries[i];
 592
 593                 if (!i) {
 594                         prev = this;
 595                         continue;
 596                 }
 597
 598                 if (prev->tsc > this->tsc) {
 599                         uint64_t diff = prev->tsc - this->tsc;
 600
 601                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 602                                                 (unsigned long long) diff);
 603                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 604                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 605                         failed++;
 606                 }
 607
 608                 prev = this;
 609         }
 610
 611         if (failed)
 612                 log_info("cs: Failed: %lu\n", failed);
 613         else
 614                 log_info("cs: Pass!\n");
 615
 616 err:
 617         free(entries);
 618         return !!failed;
 619 }
 620
 621 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 622
 623 int fio_monotonic_clocktest(void)
 624 {
 625         log_info("cs: current platform does not support CPU clocks\n");
 626         return 0;
 627 }
 628
 629 #endif