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 #ifdef ARCH_HAVE_CPU_CLOCK
  17 static unsigned long cycles_per_usec;
  18 static unsigned long last_cycles;
  19 int tsc_reliable = 0;
  20 #endif
  21 static struct timeval last_tv;
  22 static int last_tv_valid;
  23
  24 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
  25 int fio_clock_source_set = 0;
  26 enum fio_cs fio_clock_source_inited = CS_INVAL;
  27
  28 #ifdef FIO_DEBUG_TIME
  29
  30 #define HASH_BITS       8
  31 #define HASH_SIZE       (1 << HASH_BITS)
  32
  33 static struct flist_head hash[HASH_SIZE];
  34 static int gtod_inited;
  35
  36 struct gtod_log {
  37         struct flist_head list;
  38         void *caller;
  39         unsigned long calls;
  40 };
  41
  42 static struct gtod_log *find_hash(void *caller)
  43 {
  44         unsigned long h = hash_ptr(caller, HASH_BITS);
  45         struct flist_head *entry;
  46
  47         flist_for_each(entry, &hash[h]) {
  48                 struct gtod_log *log = flist_entry(entry, struct gtod_log,
  49                                                                         list);
  50
  51                 if (log->caller == caller)
  52                         return log;
  53         }
  54
  55         return NULL;
  56 }
  57
  58 static struct gtod_log *find_log(void *caller)
  59 {
  60         struct gtod_log *log = find_hash(caller);
  61
  62         if (!log) {
  63                 unsigned long h;
  64
  65                 log = malloc(sizeof(*log));
  66                 INIT_FLIST_HEAD(&log->list);
  67                 log->caller = caller;
  68                 log->calls = 0;
  69
  70                 h = hash_ptr(caller, HASH_BITS);
  71                 flist_add_tail(&log->list, &hash[h]);
  72         }
  73
  74         return log;
  75 }
  76
  77 static void gtod_log_caller(void *caller)
  78 {
  79         if (gtod_inited) {
  80                 struct gtod_log *log = find_log(caller);
  81
  82                 log->calls++;
  83         }
  84 }
  85
  86 static void fio_exit fio_dump_gtod(void)
  87 {
  88         unsigned long total_calls = 0;
  89         int i;
  90
  91         for (i = 0; i < HASH_SIZE; i++) {
  92                 struct flist_head *entry;
  93                 struct gtod_log *log;
  94
  95                 flist_for_each(entry, &hash[i]) {
  96                         log = flist_entry(entry, struct gtod_log, list);
  97
  98                         printf("function %p, calls %lu\n", log->caller,
  99                                                                 log->calls);
 100                         total_calls += log->calls;
 101                 }
 102         }
 103
 104         printf("Total %lu gettimeofday\n", total_calls);
 105 }
 106
 107 static void fio_init gtod_init(void)
 108 {
 109         int i;
 110
 111         for (i = 0; i < HASH_SIZE; i++)
 112                 INIT_FLIST_HEAD(&hash[i]);
 113
 114         gtod_inited = 1;
 115 }
 116
 117 #endif /* FIO_DEBUG_TIME */
 118
 119 #ifdef FIO_DEBUG_TIME
 120 void fio_gettime(struct timeval *tp, void *caller)
 121 #else
 122 void fio_gettime(struct timeval *tp, void fio_unused *caller)
 123 #endif
 124 {
 125 #ifdef FIO_DEBUG_TIME
 126         if (!caller)
 127                 caller = __builtin_return_address(0);
 128
 129         gtod_log_caller(caller);
 130 #endif
 131         if (fio_tv) {
 132                 memcpy(tp, fio_tv, sizeof(*tp));
 133                 return;
 134         }
 135
 136         switch (fio_clock_source) {
 137         case CS_GTOD:
 138                 gettimeofday(tp, NULL);
 139                 break;
 140         case CS_CGETTIME: {
 141                 struct timespec ts;
 142
 143 #ifdef FIO_HAVE_CLOCK_MONOTONIC
 144                 if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
 145 #else
 146                 if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
 147 #endif
 148                         log_err("fio: clock_gettime fails\n");
 149                         assert(0);
 150                 }
 151
 152                 tp->tv_sec = ts.tv_sec;
 153                 tp->tv_usec = ts.tv_nsec / 1000;
 154                 break;
 155                 }
 156 #ifdef ARCH_HAVE_CPU_CLOCK
 157         case CS_CPUCLOCK: {
 158                 unsigned long long usecs, t;
 159
 160                 t = get_cpu_clock();
 161                 if (t < last_cycles) {
 162                         dprint(FD_TIME, "CPU clock going back in time\n");
 163                         t = last_cycles;
 164                 }
 165
 166                 usecs = t / cycles_per_usec;
 167                 tp->tv_sec = usecs / 1000000;
 168                 tp->tv_usec = usecs % 1000000;
 169                 last_cycles = t;
 170                 break;
 171                 }
 172 #endif
 173         default:
 174                 log_err("fio: invalid clock source %d\n", fio_clock_source);
 175                 break;
 176         }
 177
 178         /*
 179          * If Linux is using the tsc clock on non-synced processors,
 180          * sometimes time can appear to drift backwards. Fix that up.
 181          */
 182         if (last_tv_valid) {
 183                 if (tp->tv_sec < last_tv.tv_sec)
 184                         tp->tv_sec = last_tv.tv_sec;
 185                 else if (last_tv.tv_sec == tp->tv_sec &&
 186                          tp->tv_usec < last_tv.tv_usec)
 187                         tp->tv_usec = last_tv.tv_usec;
 188         }
 189         last_tv_valid = 1;
 190         memcpy(&last_tv, tp, sizeof(*tp));
 191 }
 192
 193 #ifdef ARCH_HAVE_CPU_CLOCK
 194 static unsigned long get_cycles_per_usec(void)
 195 {
 196         struct timeval s, e;
 197         unsigned long long c_s, c_e;
 198
 199         gettimeofday(&s, NULL);
 200         c_s = get_cpu_clock();
 201         do {
 202                 unsigned long long elapsed;
 203
 204                 gettimeofday(&e, NULL);
 205                 elapsed = utime_since(&s, &e);
 206                 if (elapsed >= 1280) {
 207                         c_e = get_cpu_clock();
 208                         break;
 209                 }
 210         } while (1);
 211
 212         return (c_e - c_s + 127) >> 7;
 213 }
 214
 215 #define NR_TIME_ITERS   50
 216
 217 static void calibrate_cpu_clock(void)
 218 {
 219         double delta, mean, S;
 220         unsigned long avg, cycles[NR_TIME_ITERS];
 221         int i, samples;
 222
 223         cycles[0] = get_cycles_per_usec();
 224         S = delta = mean = 0.0;
 225         for (i = 0; i < NR_TIME_ITERS; i++) {
 226                 cycles[i] = get_cycles_per_usec();
 227                 delta = cycles[i] - mean;
 228                 if (delta) {
 229                         mean += delta / (i + 1.0);
 230                         S += delta * (cycles[i] - mean);
 231                 }
 232         }
 233
 234         S = sqrt(S / (NR_TIME_ITERS - 1.0));
 235
 236         samples = avg = 0;
 237         for (i = 0; i < NR_TIME_ITERS; i++) {
 238                 double this = cycles[i];
 239
 240                 if ((fmax(this, mean) - fmin(this, mean)) > S)
 241                         continue;
 242                 samples++;
 243                 avg += this;
 244         }
 245
 246         S /= (double) NR_TIME_ITERS;
 247         mean /= 10.0;
 248
 249         for (i = 0; i < NR_TIME_ITERS; i++)
 250                 dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
 251
 252         avg /= samples;
 253         avg = (avg + 9) / 10;
 254         dprint(FD_TIME, "avg: %lu\n", avg);
 255         dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
 256
 257         cycles_per_usec = avg;
 258 }
 259 #else
 260 static void calibrate_cpu_clock(void)
 261 {
 262 }
 263 #endif
 264
 265 void fio_clock_init(void)
 266 {
 267         if (fio_clock_source == fio_clock_source_inited)
 268                 return;
 269
 270         last_tv_valid = 0;
 271         fio_clock_source_inited = fio_clock_source;
 272         calibrate_cpu_clock();
 273
 274         /*
 275          * If the arch sets tsc_reliable != 0, then it must be good enough
 276          * to use as THE clock source. For x86 CPUs, this means the TSC
 277          * runs at a constant rate and is synced across CPU cores.
 278          */
 279         if (tsc_reliable) {
 280                 if (!fio_clock_source_set)
 281                         fio_clock_source = CS_CPUCLOCK;
 282         } else if (fio_clock_source == CS_CPUCLOCK)
 283                 log_info("fio: clocksource=cpu may not be reliable\n");
 284 }
 285
 286 unsigned long long utime_since(struct timeval *s, struct timeval *e)
 287 {
 288         long sec, usec;
 289         unsigned long long ret;
 290
 291         sec = e->tv_sec - s->tv_sec;
 292         usec = e->tv_usec - s->tv_usec;
 293         if (sec > 0 && usec < 0) {
 294                 sec--;
 295                 usec += 1000000;
 296         }
 297
 298         /*
 299          * time warp bug on some kernels?
 300          */
 301         if (sec < 0 || (sec == 0 && usec < 0))
 302                 return 0;
 303
 304         ret = sec * 1000000ULL + usec;
 305
 306         return ret;
 307 }
 308
 309 unsigned long long utime_since_now(struct timeval *s)
 310 {
 311         struct timeval t;
 312
 313         fio_gettime(&t, NULL);
 314         return utime_since(s, &t);
 315 }
 316
 317 unsigned long mtime_since(struct timeval *s, struct timeval *e)
 318 {
 319         long sec, usec, ret;
 320
 321         sec = e->tv_sec - s->tv_sec;
 322         usec = e->tv_usec - s->tv_usec;
 323         if (sec > 0 && usec < 0) {
 324                 sec--;
 325                 usec += 1000000;
 326         }
 327
 328         if (sec < 0 || (sec == 0 && usec < 0))
 329                 return 0;
 330
 331         sec *= 1000UL;
 332         usec /= 1000UL;
 333         ret = sec + usec;
 334
 335         return ret;
 336 }
 337
 338 unsigned long mtime_since_now(struct timeval *s)
 339 {
 340         struct timeval t;
 341         void *p = __builtin_return_address(0);
 342
 343         fio_gettime(&t, p);
 344         return mtime_since(s, &t);
 345 }
 346
 347 unsigned long time_since_now(struct timeval *s)
 348 {
 349         return mtime_since_now(s) / 1000;
 350 }
 351
 352 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)
 353
 354 #define CLOCK_ENTRIES   100000
 355
 356 struct clock_entry {
 357         unsigned long seq;
 358         unsigned long tsc;
 359         unsigned long cpu;
 360 };
 361
 362 struct clock_thread {
 363         pthread_t thread;
 364         int cpu;
 365         pthread_mutex_t lock;
 366         pthread_mutex_t started;
 367         uint64_t *seq;
 368         struct clock_entry *entries;
 369 };
 370
 371 static inline uint64_t atomic64_inc_return(uint64_t *seq)
 372 {
 373         return 1 + __sync_fetch_and_add(seq, 1);
 374 }
 375
 376 static void *clock_thread_fn(void *data)
 377 {
 378         struct clock_thread *t = data;
 379         struct clock_entry *c;
 380         os_cpu_mask_t cpu_mask;
 381         int i;
 382
 383         memset(&cpu_mask, 0, sizeof(cpu_mask));
 384         fio_cpu_set(&cpu_mask, t->cpu);
 385
 386         if (fio_setaffinity(gettid(), cpu_mask) == -1) {
 387                 log_err("clock setaffinity failed\n");
 388                 return (void *) 1;
 389         }
 390
 391         pthread_mutex_lock(&t->lock);
 392         pthread_mutex_unlock(&t->started);
 393
 394         c = &t->entries[0];
 395         for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
 396                 uint64_t seq, tsc;
 397
 398                 c->cpu = t->cpu;
 399                 do {
 400                         seq = atomic64_inc_return(t->seq);
 401                         tsc = get_cpu_clock();
 402                 } while (seq != *t->seq);
 403
 404                 c->seq = seq;
 405                 c->tsc = tsc;
 406         }
 407
 408         log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
 409         return NULL;
 410 }
 411
 412 static int clock_cmp(const void *p1, const void *p2)
 413 {
 414         const struct clock_entry *c1 = p1;
 415         const struct clock_entry *c2 = p2;
 416
 417         if (c1->seq == c2->seq)
 418                 log_err("cs: bug in atomic sequence!\n");
 419
 420         return c1->seq - c2->seq;
 421 }
 422
 423 int fio_monotonic_clocktest(void)
 424 {
 425         struct clock_thread *threads;
 426         unsigned int nr_cpus = cpus_online();
 427         struct clock_entry *entries;
 428         unsigned long tentries, failed;
 429         uint64_t seq = 0;
 430         int i;
 431
 432         threads = malloc(nr_cpus * sizeof(struct clock_thread));
 433         tentries = CLOCK_ENTRIES * nr_cpus;
 434         entries = malloc(tentries * sizeof(struct clock_entry));
 435
 436         log_info("cs: Testing %u CPUs\n", nr_cpus);
 437
 438         for (i = 0; i < nr_cpus; i++) {
 439                 struct clock_thread *t = &threads[i];
 440
 441                 t->cpu = i;
 442                 t->seq = &seq;
 443                 t->entries = &entries[i * CLOCK_ENTRIES];
 444                 pthread_mutex_init(&t->lock, NULL);
 445                 pthread_mutex_init(&t->started, NULL);
 446                 pthread_mutex_lock(&t->lock);
 447                 pthread_create(&t->thread, NULL, clock_thread_fn, t);
 448         }
 449
 450         for (i = 0; i < nr_cpus; i++) {
 451                 struct clock_thread *t = &threads[i];
 452
 453                 pthread_mutex_lock(&t->started);
 454         }
 455
 456         for (i = 0; i < nr_cpus; i++) {
 457                 struct clock_thread *t = &threads[i];
 458
 459                 pthread_mutex_unlock(&t->lock);
 460         }
 461
 462         for (failed = i = 0; i < nr_cpus; i++) {
 463                 struct clock_thread *t = &threads[i];
 464                 void *ret;
 465
 466                 pthread_join(t->thread, &ret);
 467                 if (ret)
 468                         failed++;
 469         }
 470         free(threads);
 471
 472         if (failed) {
 473                 log_err("Clocksource test: %u threads failed\n", failed);
 474                 goto err;
 475         }
 476
 477         qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 478
 479         for (failed = i = 0; i < tentries; i++) {
 480                 struct clock_entry *prev, *this = &entries[i];
 481
 482                 if (!i) {
 483                         prev = this;
 484                         continue;
 485                 }
 486
 487                 if (prev->tsc > this->tsc) {
 488                         uint64_t diff = prev->tsc - this->tsc;
 489
 490                         log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
 491                         log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
 492                         log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
 493                         failed++;
 494                 }
 495
 496                 prev = this;
 497         }
 498
 499         if (failed)
 500                 log_info("cs: Failed: %lu\n", failed);
 501         else
 502                 log_info("cs: Pass!\n");
 503
 504 err:
 505         free(entries);
 506         return !!failed;
 507 }
 508
 509 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 510
 511 int fio_monotonic_clocktest(void)
 512 {
 513         log_info("cs: current platform does not support CPU clocks\n");
 514         return 0;
 515 }
 516
 517 #endif