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 #endif
  20 int tsc_reliable = 0;
  21
  22 struct tv_valid {
  23         uint64_t last_cycles;
  24         int last_tv_valid;
  25         int warned;
  26 };
  27 #ifdef CONFIG_TLS_THREAD
  28 static __thread struct tv_valid static_tv_valid;
  29 #else
  30 static pthread_key_t tv_tls_key;
  31 #endif
  32
  33 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  34 int fio_clock_source_set = 0;
  35 static enum fio_cs fio_clock_source_inited = CS_INVAL;
  36
  37 #ifdef FIO_DEBUG_TIME
  38
  39 #define HASH_BITS       8
  40 #define HASH_SIZE       (1 << HASH_BITS)
  41
  42 static struct flist_head hash[HASH_SIZE];
  43 static int gtod_inited;
  44
  45 struct gtod_log {
  46         struct flist_head list;
  47         void *caller;
  48         unsigned long calls;
  49 };
  50
  51 static struct gtod_log *find_hash(void *caller)
  52 {
  53         unsigned long h = hash_ptr(caller, HASH_BITS);
  54         struct flist_head *entry;
  55
  56         flist_for_each(entry, &hash[h]) {
  57                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  58                                                                         list);
  59
  60                 if (log->caller == caller)
  61                         return log;
  62         }
  63
  64         return NULL;
  65 }
  66
  67 static struct gtod_log *find_log(void *caller)
  68 {
  69         struct gtod_log *log = find_hash(caller);
  70
  71         if (!log) {
  72                 unsigned long h;
  73
  74                 log = malloc(sizeof(*log));
  75                 INIT_FLIST_HEAD(&log->list);
  76                 log->caller = caller;
  77                 log->calls = 0;
  78
  79                 h = hash_ptr(caller, HASH_BITS);
  80                 flist_add_tail(&log->list, &hash[h]);
  81         }
  82
  83         return log;
  84 }
  85
  86 static void gtod_log_caller(void *caller)
  87 {
  88         if (gtod_inited) {
  89                 struct gtod_log *log = find_log(caller);
  90
  91                 log->calls++;
  92         }
  93 }
  94
  95 static void fio_exit fio_dump_gtod(void)
  96 {
  97         unsigned long total_calls = 0;
  98         int i;
  99
 100         for (i = 0; i < HASH_SIZE; i++) {
 101                 struct flist_head *entry;
 102                 struct gtod_log *log;
 103
 104                 flist_for_each(entry, &hash[i]) {
 105                         log = flist_entry(entry, struct gtod_log, list);
 106
 107                         printf("function %p, calls %lu\n", log->caller,
 108                                                                 log->calls);
 109                         total_calls += log->calls;
 110                 }
 111         }
 112
 113         printf("Total %lu gettimeofday\n", total_calls);
 114 }
 115
 116 static void fio_init gtod_init(void)
 117 {
 118         int i;
 119
 120         for (i = 0; i < HASH_SIZE; i++)
 121                 INIT_FLIST_HEAD(&hash[i]);
 122
 123         gtod_inited = 1;
 124 }
 125
 126 #endif /* FIO_DEBUG_TIME */
 127
 128 #ifdef CONFIG_CLOCK_GETTIME
 129 static int fill_clock_gettime(struct timespec *ts)
 130 {
 131 #ifdef CONFIG_CLOCK_MONOTONIC
 132         return clock_gettime(CLOCK_MONOTONIC, ts);
 133 #else
 134         return clock_gettime(CLOCK_REALTIME, ts);
 135 #endif
 136 }
 137 #endif
 138
 139 static void __fio_gettime(struct timeval *tp)
 140 {
 141         struct tv_valid *tv;
 142
 143 #ifdef CONFIG_TLS_THREAD
 144         tv = &static_tv_valid;
 145 #else
 146         tv = pthread_getspecific(tv_tls_key);
 147 #endif
 148
 149         switch (fio_clock_source) {
 150 #ifdef CONFIG_GETTIMEOFDAY
 151         case CS_GTOD:
 152                 gettimeofday(tp, NULL);
 153                 break;
 154 #endif
 155 #ifdef CONFIG_CLOCK_GETTIME
 156         case CS_CGETTIME: {
 157                 struct timespec ts;
 158
 159                 if (fill_clock_gettime(&ts) < 0) {
 160                         log_err("fio: clock_gettime fails\n");
 161                         assert(0);
 162                 }
 163
 164                 tp->tv_sec = ts.tv_sec;
 165                 tp->tv_usec = ts.tv_nsec / 1000;
 166                 break;
 167                 }
 168 #endif
 169 #ifdef ARCH_HAVE_CPU_CLOCK
 170         case CS_CPUCLOCK: {
 171                 uint64_t usecs, t;
 172
 173                 t = get_cpu_clock();
 174                 if (t < tv->last_cycles && tv->last_tv_valid &&
 175                     !tv->warned) {
 176                         log_err("fio: CPU clock going back in time\n");
 177                         tv->warned = 1;
 178                 }
 179
 180                 tv->last_cycles = t;
 181                 tv->last_tv_valid = 1;
 182 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 183                 usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
 184 #else
 185                 usecs = (t * inv_cycles_per_usec) / 16777216UL;
 186 #endif
 187                 tp->tv_sec = usecs / 1000000;
 188                 tp->tv_usec = usecs % 1000000;
 189                 break;
 190                 }
 191 #endif
 192         default:
 193                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 194                 break;
 195         }
 196 }
 197
 198 #ifdef FIO_DEBUG_TIME
 199 void fio_gettime(struct timeval *tp, void *caller)
 200 #else
 201 void fio_gettime(struct timeval *tp, void fio_unused *caller)
 202 #endif
 203 {
 204 #ifdef FIO_DEBUG_TIME
 205         if (!caller)
 206                 caller = __builtin_return_address(0);
 207
 208         gtod_log_caller(caller);
 209 #endif
 210         if (fio_unlikely(fio_tv)) {
 211                 memcpy(tp, fio_tv, sizeof(*tp));
 212                 return;
 213         }
 214
 215         __fio_gettime(tp);
 216 }
 217
 218 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 219 static unsigned long get_cycles_per_usec(void)
 220 {
 221         struct timeval s, e;
 222         uint64_t c_s, c_e;
 223         enum fio_cs old_cs = fio_clock_source;
 224
 225 #ifdef CONFIG_CLOCK_GETTIME
 226         fio_clock_source = CS_CGETTIME;
 227 #else
 228         fio_clock_source = CS_GTOD;
 229 #endif
 230         __fio_gettime(&s);
 231
 232         c_s = get_cpu_clock();
 233         do {
 234                 uint64_t elapsed;
 235
 236                 __fio_gettime(&e);
 237
 238                 elapsed = utime_since(&s, &e);
 239                 if (elapsed >= 1280) {
 240                         c_e = get_cpu_clock();
 241                         break;
 242                 }
 243         } while (1);
 244
 245         fio_clock_source = old_cs;
 246         return (c_e - c_s + 127) >> 7;
 247 }
 248
 249 #define NR_TIME_ITERS   50
 250
 251 static int calibrate_cpu_clock(void)
 252 {
 253         double delta, mean, S;
 254         uint64_t avg, cycles[NR_TIME_ITERS];
 255         int i, samples;
 256
 257         cycles[0] = get_cycles_per_usec();
 258         S = delta = mean = 0.0;
 259         for (i = 0; i < NR_TIME_ITERS; i++) {
 260                 cycles[i] = get_cycles_per_usec();
 261                 delta = cycles[i] - mean;
 262                 if (delta) {
 263                         mean += delta / (i + 1.0);
 264                         S += delta * (cycles[i] - mean);
 265                 }
 266         }
 267
 268         /*
 269          * The most common platform clock breakage is returning zero
 270          * indefinitely. Check for that and return failure.
 271          */
 272         if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
 273                 return 1;
 274
 275         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 276
 277         samples = avg = 0;
 278         for (i = 0; i < NR_TIME_ITERS; i++) {
 279                 double this = cycles[i];
 280
 281                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 282                         continue;
 283                 samples++;
 284                 avg += this;
 285         }
 286
 287         S /= (double) NR_TIME_ITERS;
 288         mean /= 10.0;
 289
 290         for (i = 0; i < NR_TIME_ITERS; i++)
 291                 dprint(FD_TIME, "cycles[%d]=%llu\n", i,
 292                                         (unsigned long long) cycles[i] / 10);
 293
 294         avg /= samples;
 295         avg = (avg + 5) / 10;
 296         dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
 297         dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
 298
 299         cycles_per_usec = avg;
 300         inv_cycles_per_usec = 16777216UL / cycles_per_usec;
 301         dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
 302         return 0;
 303 }
 304 #else
 305 static int calibrate_cpu_clock(void)
 306 {
 307 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
 308         return 0;
 309 #else
 310         return 1;
 311 #endif
 312 }
 313 #endif // ARCH_HAVE_CPU_CLOCK
 314
 315 #ifndef CONFIG_TLS_THREAD
 316 void fio_local_clock_init(int is_thread)
 317 {
 318         struct tv_valid *t;
 319
 320         t = calloc(1, sizeof(*t));
 321         if (pthread_setspecific(tv_tls_key, t)) {
 322                 log_err("fio: can't set TLS key\n");
 323                 assert(0);
 324         }
 325 }
 326
 327 static void kill_tv_tls_key(void *data)
 328 {
 329         free(data);
 330 }
 331 #else
 332 void fio_local_clock_init(int is_thread)
 333 {
 334 }
 335 #endif
 336
 337 void fio_clock_init(void)
 338 {
 339         if (fio_clock_source == fio_clock_source_inited)
 340                 return;
 341
 342 #ifndef CONFIG_TLS_THREAD
 343         if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
 344                 log_err("fio: can't create TLS key\n");
 345 #endif
 346
 347         fio_clock_source_inited = fio_clock_source;
 348
 349         if (calibrate_cpu_clock())
 350                 tsc_reliable = 0;
 351
 352         /*
 353          * If the arch sets tsc_reliable != 0, then it must be good enough
 354          * to use as THE clock source. For x86 CPUs, this means the TSC
 355          * runs at a constant rate and is synced across CPU cores.
 356          */
 357         if (tsc_reliable) {
 358                 if (!fio_clock_source_set)
 359                         fio_clock_source = CS_CPUCLOCK;
 360         } else if (fio_clock_source == CS_CPUCLOCK)
 361                 log_info("fio: clocksource=cpu may not be reliable\n");
 362 }
 363
 364 uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 365 {
 366         long sec, usec;
 367         uint64_t ret;
 368
 369         sec = e->tv_sec - s->tv_sec;
 370         usec = e->tv_usec - s->tv_usec;
 371         if (sec > 0 && usec < 0) {
 372                 sec--;
 373                 usec += 1000000;
 374         }
 375
 376         /*
 377          * time warp bug on some kernels?
 378          */
 379         if (sec < 0 || (sec == 0 && usec < 0))
 380                 return 0;
 381
 382         ret = sec * 1000000ULL + usec;
 383
 384         return ret;
 385 }
 386
 387 uint64_t utime_since_now(const struct timeval *s)
 388 {
 389         struct timeval t;
 390
 391         fio_gettime(&t, NULL);
 392         return utime_since(s, &t);
 393 }
 394
 395 uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
 396 {
 397         long sec, usec, ret;
 398
 399         sec = e->tv_sec - s->tv_sec;
 400         usec = e->tv_usec - s->tv_usec;
 401         if (sec > 0 && usec < 0) {
 402                 sec--;
 403                 usec += 1000000;
 404         }
 405
 406         if (sec < 0 || (sec == 0 && usec < 0))
 407                 return 0;
 408
 409         sec *= 1000UL;
 410         usec /= 1000UL;
 411         ret = sec + usec;
 412
 413         return ret;
 414 }
 415
 416 uint64_t mtime_since_now(const struct timeval *s)
 417 {
 418         struct timeval t;
 419         void *p = __builtin_return_address(0);
 420
 421         fio_gettime(&t, p);
 422         return mtime_since(s, &t);
 423 }
 424
 425 uint64_t time_since_now(const struct timeval *s)
 426 {
 427         return mtime_since_now(s) / 1000;
 428 }
 429
 430 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
 431     defined(CONFIG_SFAA)
 432
 433 #define CLOCK_ENTRIES   100000
 434
 435 struct clock_entry {
 436         uint32_t seq;
 437         uint32_t cpu;
 438         uint64_t tsc;
 439 };
 440
 441 struct clock_thread {
 442         pthread_t thread;
 443         int cpu;
 444         pthread_mutex_t lock;
 445         pthread_mutex_t started;
 446         uint32_t *seq;
 447         struct clock_entry *entries;
 448 };
 449
 450 static inline uint32_t atomic32_inc_return(uint32_t *seq)
 451 {
 452         return 1 + __sync_fetch_and_add(seq, 1);
 453 }
 454
 455 static void *clock_thread_fn(void *data)
 456 {
 457         struct clock_thread *t = data;
 458         struct clock_entry *c;
 459         os_cpu_mask_t cpu_mask;
 460         uint32_t last_seq;
 461         int i;
 462
 463         memset(&cpu_mask, 0, sizeof(cpu_mask));
 464         fio_cpu_set(&cpu_mask, t->cpu);
 465
 466         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 467                 log_err("clock setaffinity failed\n");
 468                 return (void *) 1;
 469         }
 470
 471         pthread_mutex_lock(&t->lock);
 472         pthread_mutex_unlock(&t->started);
 473
 474         last_seq = 0;
 475         c = &t->entries[0];
 476         for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
 477                 uint32_t seq;
 478                 uint64_t tsc;
 479
 480                 c->cpu = t->cpu;
 481                 do {
 482                         seq = atomic32_inc_return(t->seq);
 483                         if (seq < last_seq)
 484                                 break;
 485                         tsc = get_cpu_clock();
 486                 } while (seq != *t->seq);
 487
 488                 c->seq = seq;
 489                 c->tsc = tsc;
 490         }
 491
 492         log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
 493                 (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
 494
 495         /*
 496          * The most common platform clock breakage is returning zero
 497          * indefinitely. Check for that and return failure.
 498          */
 499         if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
 500                 return (void *) 1;
 501
 502         return NULL;
 503 }
 504
 505 static int clock_cmp(const void *p1, const void *p2)
 506 {
 507         const struct clock_entry *c1 = p1;
 508         const struct clock_entry *c2 = p2;
 509
 510         if (c1->seq == c2->seq)
 511                 log_err("cs: bug in atomic sequence!\n");
 512
 513         return c1->seq - c2->seq;
 514 }
 515
 516 int fio_monotonic_clocktest(void)
 517 {
 518         struct clock_thread *cthreads;
 519         unsigned int nr_cpus = cpus_online();
 520         struct clock_entry *entries;
 521         unsigned long tentries, failed = 0;
 522         struct clock_entry *prev, *this;
 523         uint32_t seq = 0;
 524         unsigned int i;
 525
 526         log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 527
 528 #ifdef FIO_INC_DEBUG
 529         fio_debug |= 1U << FD_TIME;
 530 #endif
 531         calibrate_cpu_clock();
 532 #ifdef FIO_INC_DEBUG
 533         fio_debug &= ~(1U << FD_TIME);
 534 #endif
 535
 536         cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
 537         tentries = CLOCK_ENTRIES * nr_cpus;
 538         entries = malloc(tentries * sizeof(struct clock_entry));
 539
 540         log_info("cs: Testing %u CPUs\n", nr_cpus);
 541
 542         for (i = 0; i < nr_cpus; i++) {
 543                 struct clock_thread *t = &cthreads[i];
 544
 545                 t->cpu = i;
 546                 t->seq = &seq;
 547                 t->entries = &entries[i * CLOCK_ENTRIES];
 548                 pthread_mutex_init(&t->lock, NULL);
 549                 pthread_mutex_init(&t->started, NULL);
 550                 pthread_mutex_lock(&t->lock);
 551                 if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
 552                         failed++;
 553                         nr_cpus = i;
 554                         break;
 555                 }
 556         }
 557
 558         for (i = 0; i < nr_cpus; i++) {
 559                 struct clock_thread *t = &cthreads[i];
 560
 561                 pthread_mutex_lock(&t->started);
 562         }
 563
 564         for (i = 0; i < nr_cpus; i++) {
 565                 struct clock_thread *t = &cthreads[i];
 566
 567                 pthread_mutex_unlock(&t->lock);
 568         }
 569
 570         for (i = 0; i < nr_cpus; i++) {
 571                 struct clock_thread *t = &cthreads[i];
 572                 void *ret;
 573
 574                 pthread_join(t->thread, &ret);
 575                 if (ret)
 576                         failed++;
 577         }
 578         free(cthreads);
 579
 580         if (failed) {
 581                 log_err("Clocksource test: %lu threads failed\n", failed);
 582                 goto err;
 583         }
 584
 585         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 586
 587         for (failed = i = 0; i < tentries; i++) {
 588                 this = &entries[i];
 589
 590                 if (!i) {
 591                         prev = this;
 592                         continue;
 593                 }
 594
 595                 if (prev->tsc > this->tsc) {
 596                         uint64_t diff = prev->tsc - this->tsc;
 597
 598                         log_info("cs: CPU clock mismatch (diff=%llu):\n",
 599                                                 (unsigned long long) diff);
 600                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
 601                         log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
 602                         failed++;
 603                 }
 604
 605                 prev = this;
 606         }
 607
 608         if (failed)
 609                 log_info("cs: Failed: %lu\n", failed);
 610         else
 611                 log_info("cs: Pass!\n");
 612
 613 err:
 614         free(entries);
 615         return !!failed;
 616 }
 617
 618 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 619
 620 int fio_monotonic_clocktest(void)
 621 {
 622         log_info("cs: current platform does not support CPU clocks\n");
 623         return 0;
 624 }
 625
 626 #endif