| 1 | /* |
| 2 | * Clock functions |
| 3 | */ |
| 4 | |
| 5 | #include <math.h> |
| 6 | |
| 7 | #include "fio.h" |
| 8 | #include "os/os.h" |
| 9 | |
| 10 | #if defined(ARCH_HAVE_CPU_CLOCK) |
| 11 | #ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC |
| 12 | static unsigned long long cycles_per_msec; |
| 13 | static unsigned long long cycles_start; |
| 14 | static unsigned long long clock_mult; |
| 15 | static unsigned long long max_cycles_mask; |
| 16 | static unsigned long long nsecs_for_max_cycles; |
| 17 | static unsigned int clock_shift; |
| 18 | static unsigned int max_cycles_shift; |
| 19 | #define MAX_CLOCK_SEC 60*60 |
| 20 | #endif |
| 21 | #ifdef ARCH_CPU_CLOCK_WRAPS |
| 22 | static unsigned int cycles_wrap; |
| 23 | #endif |
| 24 | #endif |
| 25 | bool tsc_reliable = false; |
| 26 | |
| 27 | struct tv_valid { |
| 28 | int warned; |
| 29 | }; |
| 30 | #ifdef ARCH_HAVE_CPU_CLOCK |
| 31 | #ifdef CONFIG_TLS_THREAD |
| 32 | static __thread struct tv_valid static_tv_valid; |
| 33 | #else |
| 34 | static pthread_key_t tv_tls_key; |
| 35 | #endif |
| 36 | #endif |
| 37 | |
| 38 | enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE; |
| 39 | int fio_clock_source_set = 0; |
| 40 | static enum fio_cs fio_clock_source_inited = CS_INVAL; |
| 41 | |
| 42 | #ifdef FIO_DEBUG_TIME |
| 43 | |
| 44 | #define HASH_BITS 8 |
| 45 | #define HASH_SIZE (1 << HASH_BITS) |
| 46 | |
| 47 | static struct flist_head hash[HASH_SIZE]; |
| 48 | static int gtod_inited; |
| 49 | |
| 50 | struct gtod_log { |
| 51 | struct flist_head list; |
| 52 | void *caller; |
| 53 | unsigned long calls; |
| 54 | }; |
| 55 | |
| 56 | static struct gtod_log *find_hash(void *caller) |
| 57 | { |
| 58 | unsigned long h = hash_ptr(caller, HASH_BITS); |
| 59 | struct flist_head *entry; |
| 60 | |
| 61 | flist_for_each(entry, &hash[h]) { |
| 62 | struct gtod_log *log = flist_entry(entry, struct gtod_log, |
| 63 | list); |
| 64 | |
| 65 | if (log->caller == caller) |
| 66 | return log; |
| 67 | } |
| 68 | |
| 69 | return NULL; |
| 70 | } |
| 71 | |
| 72 | static void inc_caller(void *caller) |
| 73 | { |
| 74 | struct gtod_log *log = find_hash(caller); |
| 75 | |
| 76 | if (!log) { |
| 77 | unsigned long h; |
| 78 | |
| 79 | log = malloc(sizeof(*log)); |
| 80 | INIT_FLIST_HEAD(&log->list); |
| 81 | log->caller = caller; |
| 82 | log->calls = 0; |
| 83 | |
| 84 | h = hash_ptr(caller, HASH_BITS); |
| 85 | flist_add_tail(&log->list, &hash[h]); |
| 86 | } |
| 87 | |
| 88 | log->calls++; |
| 89 | } |
| 90 | |
| 91 | static void gtod_log_caller(void *caller) |
| 92 | { |
| 93 | if (gtod_inited) |
| 94 | inc_caller(caller); |
| 95 | } |
| 96 | |
| 97 | static void fio_exit fio_dump_gtod(void) |
| 98 | { |
| 99 | unsigned long total_calls = 0; |
| 100 | int i; |
| 101 | |
| 102 | for (i = 0; i < HASH_SIZE; i++) { |
| 103 | struct flist_head *entry; |
| 104 | struct gtod_log *log; |
| 105 | |
| 106 | flist_for_each(entry, &hash[i]) { |
| 107 | log = flist_entry(entry, struct gtod_log, list); |
| 108 | |
| 109 | printf("function %p, calls %lu\n", log->caller, |
| 110 | log->calls); |
| 111 | total_calls += log->calls; |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | printf("Total %lu gettimeofday\n", total_calls); |
| 116 | } |
| 117 | |
| 118 | static void fio_init gtod_init(void) |
| 119 | { |
| 120 | int i; |
| 121 | |
| 122 | for (i = 0; i < HASH_SIZE; i++) |
| 123 | INIT_FLIST_HEAD(&hash[i]); |
| 124 | |
| 125 | gtod_inited = 1; |
| 126 | } |
| 127 | |
| 128 | #endif /* FIO_DEBUG_TIME */ |
| 129 | |
| 130 | /* |
| 131 | * Queries the value of the monotonic clock if a monotonic clock is available |
| 132 | * or the wall clock time if no monotonic clock is available. Returns 0 if |
| 133 | * querying the clock succeeded or -1 if querying the clock failed. |
| 134 | */ |
| 135 | int fio_get_mono_time(struct timespec *ts) |
| 136 | { |
| 137 | int ret; |
| 138 | |
| 139 | #ifdef CONFIG_CLOCK_GETTIME |
| 140 | #if defined(CONFIG_CLOCK_MONOTONIC) |
| 141 | ret = clock_gettime(CLOCK_MONOTONIC, ts); |
| 142 | #else |
| 143 | ret = clock_gettime(CLOCK_REALTIME, ts); |
| 144 | #endif |
| 145 | #else |
| 146 | struct timeval tv; |
| 147 | |
| 148 | ret = gettimeofday(&tv, NULL); |
| 149 | if (ret == 0) { |
| 150 | ts->tv_sec = tv.tv_sec; |
| 151 | ts->tv_nsec = tv.tv_usec * 1000; |
| 152 | } |
| 153 | #endif |
| 154 | assert(ret <= 0); |
| 155 | return ret; |
| 156 | } |
| 157 | |
| 158 | static void __fio_gettime(struct timespec *tp) |
| 159 | { |
| 160 | switch (fio_clock_source) { |
| 161 | #ifdef CONFIG_GETTIMEOFDAY |
| 162 | case CS_GTOD: { |
| 163 | struct timeval tv; |
| 164 | gettimeofday(&tv, NULL); |
| 165 | |
| 166 | tp->tv_sec = tv.tv_sec; |
| 167 | tp->tv_nsec = tv.tv_usec * 1000; |
| 168 | break; |
| 169 | } |
| 170 | #endif |
| 171 | #ifdef CONFIG_CLOCK_GETTIME |
| 172 | case CS_CGETTIME: { |
| 173 | if (fio_get_mono_time(tp) < 0) { |
| 174 | log_err("fio: fio_get_mono_time() fails\n"); |
| 175 | assert(0); |
| 176 | } |
| 177 | break; |
| 178 | } |
| 179 | #endif |
| 180 | #ifdef ARCH_HAVE_CPU_CLOCK |
| 181 | case CS_CPUCLOCK: { |
| 182 | uint64_t nsecs, t, multiples; |
| 183 | struct tv_valid *tv; |
| 184 | |
| 185 | #ifdef CONFIG_TLS_THREAD |
| 186 | tv = &static_tv_valid; |
| 187 | #else |
| 188 | tv = pthread_getspecific(tv_tls_key); |
| 189 | #endif |
| 190 | |
| 191 | t = get_cpu_clock(); |
| 192 | #ifdef ARCH_CPU_CLOCK_WRAPS |
| 193 | if (t < cycles_start && !cycles_wrap) |
| 194 | cycles_wrap = 1; |
| 195 | else if (cycles_wrap && t >= cycles_start && !tv->warned) { |
| 196 | log_err("fio: double CPU clock wrap\n"); |
| 197 | tv->warned = 1; |
| 198 | } |
| 199 | #endif |
| 200 | #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC |
| 201 | nsecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC * 1000; |
| 202 | #else |
| 203 | t -= cycles_start; |
| 204 | multiples = t >> max_cycles_shift; |
| 205 | nsecs = multiples * nsecs_for_max_cycles; |
| 206 | nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift; |
| 207 | #endif |
| 208 | tp->tv_sec = nsecs / 1000000000ULL; |
| 209 | tp->tv_nsec = nsecs % 1000000000ULL; |
| 210 | break; |
| 211 | } |
| 212 | #endif |
| 213 | default: |
| 214 | log_err("fio: invalid clock source %d\n", fio_clock_source); |
| 215 | break; |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | #ifdef FIO_DEBUG_TIME |
| 220 | void fio_gettime(struct timespec *tp, void *caller) |
| 221 | #else |
| 222 | void fio_gettime(struct timespec *tp, void fio_unused *caller) |
| 223 | #endif |
| 224 | { |
| 225 | #ifdef FIO_DEBUG_TIME |
| 226 | if (!caller) |
| 227 | caller = __builtin_return_address(0); |
| 228 | |
| 229 | gtod_log_caller(caller); |
| 230 | #endif |
| 231 | if (fio_unlikely(fio_gettime_offload(tp))) |
| 232 | return; |
| 233 | |
| 234 | __fio_gettime(tp); |
| 235 | } |
| 236 | |
| 237 | #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC) |
| 238 | static unsigned long get_cycles_per_msec(void) |
| 239 | { |
| 240 | struct timespec s, e; |
| 241 | uint64_t c_s, c_e; |
| 242 | uint64_t elapsed; |
| 243 | |
| 244 | fio_get_mono_time(&s); |
| 245 | |
| 246 | c_s = get_cpu_clock(); |
| 247 | do { |
| 248 | fio_get_mono_time(&e); |
| 249 | c_e = get_cpu_clock(); |
| 250 | |
| 251 | elapsed = ntime_since(&s, &e); |
| 252 | if (elapsed >= 1280000) |
| 253 | break; |
| 254 | } while (1); |
| 255 | |
| 256 | return (c_e - c_s) * 1000000 / 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, sft = 0; |
| 266 | unsigned long long tmp, max_ticks, max_mult; |
| 267 | |
| 268 | cycles[0] = get_cycles_per_msec(); |
| 269 | S = delta = mean = 0.0; |
| 270 | for (i = 0; i < NR_TIME_ITERS; i++) { |
| 271 | cycles[i] = get_cycles_per_msec(); |
| 272 | delta = cycles[i] - mean; |
| 273 | if (delta) { |
| 274 | mean += delta / (i + 1.0); |
| 275 | S += delta * (cycles[i] - mean); |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * The most common platform clock breakage is returning zero |
| 281 | * indefinitely. Check for that and return failure. |
| 282 | */ |
| 283 | if (!cycles[0] && !cycles[NR_TIME_ITERS - 1]) |
| 284 | return 1; |
| 285 | |
| 286 | S = sqrt(S / (NR_TIME_ITERS - 1.0)); |
| 287 | |
| 288 | minc = -1ULL; |
| 289 | maxc = samples = avg = 0; |
| 290 | for (i = 0; i < NR_TIME_ITERS; i++) { |
| 291 | double this = cycles[i]; |
| 292 | |
| 293 | minc = min(cycles[i], minc); |
| 294 | maxc = max(cycles[i], maxc); |
| 295 | |
| 296 | if ((fmax(this, mean) - fmin(this, mean)) > S) |
| 297 | continue; |
| 298 | samples++; |
| 299 | avg += this; |
| 300 | } |
| 301 | |
| 302 | S /= (double) NR_TIME_ITERS; |
| 303 | |
| 304 | for (i = 0; i < NR_TIME_ITERS; i++) |
| 305 | dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]); |
| 306 | |
| 307 | avg /= samples; |
| 308 | cycles_per_msec = avg; |
| 309 | dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f, N=%d\n", |
| 310 | (unsigned long long) minc, |
| 311 | (unsigned long long) maxc, mean, S, NR_TIME_ITERS); |
| 312 | dprint(FD_TIME, "trimmed mean=%llu, N=%d\n", (unsigned long long) avg, samples); |
| 313 | |
| 314 | max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL; |
| 315 | max_mult = ULLONG_MAX / max_ticks; |
| 316 | dprint(FD_TIME, "max_ticks=%llu, __builtin_clzll=%d, " |
| 317 | "max_mult=%llu\n", max_ticks, |
| 318 | __builtin_clzll(max_ticks), max_mult); |
| 319 | |
| 320 | /* |
| 321 | * Find the largest shift count that will produce |
| 322 | * a multiplier that does not exceed max_mult |
| 323 | */ |
| 324 | tmp = max_mult * cycles_per_msec / 1000000; |
| 325 | while (tmp > 1) { |
| 326 | tmp >>= 1; |
| 327 | sft++; |
| 328 | dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft); |
| 329 | } |
| 330 | |
| 331 | clock_shift = sft; |
| 332 | clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec; |
| 333 | dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift, |
| 334 | clock_mult); |
| 335 | |
| 336 | /* |
| 337 | * Find the greatest power of 2 clock ticks that is less than the |
| 338 | * ticks in MAX_CLOCK_SEC |
| 339 | */ |
| 340 | max_cycles_shift = max_cycles_mask = 0; |
| 341 | tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec; |
| 342 | dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, |
| 343 | max_cycles_shift); |
| 344 | while (tmp > 1) { |
| 345 | tmp >>= 1; |
| 346 | max_cycles_shift++; |
| 347 | dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift); |
| 348 | } |
| 349 | /* |
| 350 | * if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec |
| 351 | * here we will have a discontinuity every |
| 352 | * (1ULL << max_cycles_shift) cycles |
| 353 | */ |
| 354 | nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult) |
| 355 | >> clock_shift; |
| 356 | |
| 357 | /* Use a bitmask to calculate ticks % (1ULL << max_cycles_shift) */ |
| 358 | for (tmp = 0; tmp < max_cycles_shift; tmp++) |
| 359 | max_cycles_mask |= 1ULL << tmp; |
| 360 | |
| 361 | dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, " |
| 362 | "nsecs_for_max_cycles=%llu, " |
| 363 | "max_cycles_mask=%016llx\n", |
| 364 | max_cycles_shift, (1ULL << max_cycles_shift), |
| 365 | nsecs_for_max_cycles, max_cycles_mask); |
| 366 | |
| 367 | cycles_start = get_cpu_clock(); |
| 368 | dprint(FD_TIME, "cycles_start=%llu\n", cycles_start); |
| 369 | return 0; |
| 370 | } |
| 371 | #else |
| 372 | static int calibrate_cpu_clock(void) |
| 373 | { |
| 374 | #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC |
| 375 | return 0; |
| 376 | #else |
| 377 | return 1; |
| 378 | #endif |
| 379 | } |
| 380 | #endif // ARCH_HAVE_CPU_CLOCK |
| 381 | |
| 382 | #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(CONFIG_TLS_THREAD) |
| 383 | void fio_local_clock_init(void) |
| 384 | { |
| 385 | struct tv_valid *t; |
| 386 | |
| 387 | t = calloc(1, sizeof(*t)); |
| 388 | if (pthread_setspecific(tv_tls_key, t)) { |
| 389 | log_err("fio: can't set TLS key\n"); |
| 390 | assert(0); |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | static void kill_tv_tls_key(void *data) |
| 395 | { |
| 396 | free(data); |
| 397 | } |
| 398 | #else |
| 399 | void fio_local_clock_init(void) |
| 400 | { |
| 401 | } |
| 402 | #endif |
| 403 | |
| 404 | void fio_clock_init(void) |
| 405 | { |
| 406 | if (fio_clock_source == fio_clock_source_inited) |
| 407 | return; |
| 408 | |
| 409 | #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(CONFIG_TLS_THREAD) |
| 410 | if (pthread_key_create(&tv_tls_key, kill_tv_tls_key)) |
| 411 | log_err("fio: can't create TLS key\n"); |
| 412 | #endif |
| 413 | |
| 414 | fio_clock_source_inited = fio_clock_source; |
| 415 | |
| 416 | if (calibrate_cpu_clock()) |
| 417 | tsc_reliable = false; |
| 418 | |
| 419 | /* |
| 420 | * If the arch sets tsc_reliable != 0, then it must be good enough |
| 421 | * to use as THE clock source. For x86 CPUs, this means the TSC |
| 422 | * runs at a constant rate and is synced across CPU cores. |
| 423 | */ |
| 424 | if (tsc_reliable) { |
| 425 | if (!fio_clock_source_set && !fio_monotonic_clocktest(0)) |
| 426 | fio_clock_source = CS_CPUCLOCK; |
| 427 | } else if (fio_clock_source == CS_CPUCLOCK) |
| 428 | log_info("fio: clocksource=cpu may not be reliable\n"); |
| 429 | dprint(FD_TIME, "gettime: clocksource=%d\n", (int) fio_clock_source); |
| 430 | } |
| 431 | |
| 432 | uint64_t ntime_since(const struct timespec *s, const struct timespec *e) |
| 433 | { |
| 434 | int64_t sec, nsec; |
| 435 | |
| 436 | sec = e->tv_sec - s->tv_sec; |
| 437 | nsec = e->tv_nsec - s->tv_nsec; |
| 438 | if (sec > 0 && nsec < 0) { |
| 439 | sec--; |
| 440 | nsec += 1000000000LL; |
| 441 | } |
| 442 | |
| 443 | /* |
| 444 | * time warp bug on some kernels? |
| 445 | */ |
| 446 | if (sec < 0 || (sec == 0 && nsec < 0)) |
| 447 | return 0; |
| 448 | |
| 449 | return nsec + (sec * 1000000000LL); |
| 450 | } |
| 451 | |
| 452 | uint64_t ntime_since_now(const struct timespec *s) |
| 453 | { |
| 454 | struct timespec now; |
| 455 | |
| 456 | fio_gettime(&now, NULL); |
| 457 | return ntime_since(s, &now); |
| 458 | } |
| 459 | |
| 460 | uint64_t utime_since(const struct timespec *s, const struct timespec *e) |
| 461 | { |
| 462 | int64_t sec, usec; |
| 463 | |
| 464 | sec = e->tv_sec - s->tv_sec; |
| 465 | usec = (e->tv_nsec - s->tv_nsec) / 1000; |
| 466 | if (sec > 0 && usec < 0) { |
| 467 | sec--; |
| 468 | usec += 1000000; |
| 469 | } |
| 470 | |
| 471 | /* |
| 472 | * time warp bug on some kernels? |
| 473 | */ |
| 474 | if (sec < 0 || (sec == 0 && usec < 0)) |
| 475 | return 0; |
| 476 | |
| 477 | return usec + (sec * 1000000); |
| 478 | } |
| 479 | |
| 480 | uint64_t utime_since_now(const struct timespec *s) |
| 481 | { |
| 482 | struct timespec t; |
| 483 | #ifdef FIO_DEBUG_TIME |
| 484 | void *p = __builtin_return_address(0); |
| 485 | |
| 486 | fio_gettime(&t, p); |
| 487 | #else |
| 488 | fio_gettime(&t, NULL); |
| 489 | #endif |
| 490 | |
| 491 | return utime_since(s, &t); |
| 492 | } |
| 493 | |
| 494 | uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e) |
| 495 | { |
| 496 | int64_t sec, usec; |
| 497 | |
| 498 | sec = e->tv_sec - s->tv_sec; |
| 499 | usec = (e->tv_usec - s->tv_usec); |
| 500 | if (sec > 0 && usec < 0) { |
| 501 | sec--; |
| 502 | usec += 1000000; |
| 503 | } |
| 504 | |
| 505 | if (sec < 0 || (sec == 0 && usec < 0)) |
| 506 | return 0; |
| 507 | |
| 508 | sec *= 1000; |
| 509 | usec /= 1000; |
| 510 | return sec + usec; |
| 511 | } |
| 512 | |
| 513 | uint64_t mtime_since_now(const struct timespec *s) |
| 514 | { |
| 515 | struct timespec t; |
| 516 | #ifdef FIO_DEBUG_TIME |
| 517 | void *p = __builtin_return_address(0); |
| 518 | |
| 519 | fio_gettime(&t, p); |
| 520 | #else |
| 521 | fio_gettime(&t, NULL); |
| 522 | #endif |
| 523 | |
| 524 | return mtime_since(s, &t); |
| 525 | } |
| 526 | |
| 527 | /* |
| 528 | * Returns *e - *s in milliseconds as a signed integer. Note: rounding is |
| 529 | * asymmetric. If the difference yields +1 ns then 0 is returned. If the |
| 530 | * difference yields -1 ns then -1 is returned. |
| 531 | */ |
| 532 | int64_t rel_time_since(const struct timespec *s, const struct timespec *e) |
| 533 | { |
| 534 | int64_t sec, nsec; |
| 535 | |
| 536 | sec = e->tv_sec - s->tv_sec; |
| 537 | nsec = e->tv_nsec - s->tv_nsec; |
| 538 | if (nsec < 0) { |
| 539 | sec--; |
| 540 | nsec += 1000ULL * 1000 * 1000; |
| 541 | } |
| 542 | assert(0 <= nsec && nsec < 1000ULL * 1000 * 1000); |
| 543 | |
| 544 | return sec * 1000 + nsec / (1000 * 1000); |
| 545 | } |
| 546 | |
| 547 | /* |
| 548 | * Returns *e - *s in milliseconds as an unsigned integer. Returns 0 if |
| 549 | * *e < *s. |
| 550 | */ |
| 551 | uint64_t mtime_since(const struct timespec *s, const struct timespec *e) |
| 552 | { |
| 553 | return max(rel_time_since(s, e), (int64_t)0); |
| 554 | } |
| 555 | |
| 556 | uint64_t time_since_now(const struct timespec *s) |
| 557 | { |
| 558 | return mtime_since_now(s) / 1000; |
| 559 | } |
| 560 | |
| 561 | #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \ |
| 562 | defined(CONFIG_SYNC_SYNC) && defined(CONFIG_CMP_SWAP) |
| 563 | |
| 564 | #define CLOCK_ENTRIES_DEBUG 100000 |
| 565 | #define CLOCK_ENTRIES_TEST 1000 |
| 566 | |
| 567 | struct clock_entry { |
| 568 | uint32_t seq; |
| 569 | uint32_t cpu; |
| 570 | uint64_t tsc; |
| 571 | }; |
| 572 | |
| 573 | struct clock_thread { |
| 574 | pthread_t thread; |
| 575 | int cpu; |
| 576 | int debug; |
| 577 | struct fio_sem lock; |
| 578 | unsigned long nr_entries; |
| 579 | uint32_t *seq; |
| 580 | struct clock_entry *entries; |
| 581 | }; |
| 582 | |
| 583 | static inline uint32_t atomic32_compare_and_swap(uint32_t *ptr, uint32_t old, |
| 584 | uint32_t new) |
| 585 | { |
| 586 | return __sync_val_compare_and_swap(ptr, old, new); |
| 587 | } |
| 588 | |
| 589 | static void *clock_thread_fn(void *data) |
| 590 | { |
| 591 | struct clock_thread *t = data; |
| 592 | struct clock_entry *c; |
| 593 | os_cpu_mask_t cpu_mask; |
| 594 | unsigned long long first; |
| 595 | int i; |
| 596 | |
| 597 | if (fio_cpuset_init(&cpu_mask)) { |
| 598 | int __err = errno; |
| 599 | |
| 600 | log_err("clock cpuset init failed: %s\n", strerror(__err)); |
| 601 | goto err_out; |
| 602 | } |
| 603 | |
| 604 | fio_cpu_set(&cpu_mask, t->cpu); |
| 605 | |
| 606 | if (fio_setaffinity(gettid(), cpu_mask) == -1) { |
| 607 | int __err = errno; |
| 608 | |
| 609 | log_err("clock setaffinity failed: %s\n", strerror(__err)); |
| 610 | goto err; |
| 611 | } |
| 612 | |
| 613 | fio_sem_down(&t->lock); |
| 614 | |
| 615 | first = get_cpu_clock(); |
| 616 | c = &t->entries[0]; |
| 617 | for (i = 0; i < t->nr_entries; i++, c++) { |
| 618 | uint32_t seq; |
| 619 | uint64_t tsc; |
| 620 | |
| 621 | c->cpu = t->cpu; |
| 622 | do { |
| 623 | seq = *t->seq; |
| 624 | if (seq == UINT_MAX) |
| 625 | break; |
| 626 | __sync_synchronize(); |
| 627 | tsc = get_cpu_clock(); |
| 628 | } while (seq != atomic32_compare_and_swap(t->seq, seq, seq + 1)); |
| 629 | |
| 630 | if (seq == UINT_MAX) |
| 631 | break; |
| 632 | |
| 633 | c->seq = seq; |
| 634 | c->tsc = tsc; |
| 635 | } |
| 636 | |
| 637 | if (t->debug) { |
| 638 | unsigned long long clocks; |
| 639 | |
| 640 | clocks = t->entries[i - 1].tsc - t->entries[0].tsc; |
| 641 | log_info("cs: cpu%3d: %llu clocks seen, first %llu\n", t->cpu, |
| 642 | clocks, first); |
| 643 | } |
| 644 | |
| 645 | /* |
| 646 | * The most common platform clock breakage is returning zero |
| 647 | * indefinitely. Check for that and return failure. |
| 648 | */ |
| 649 | if (i > 1 && !t->entries[i - 1].tsc && !t->entries[0].tsc) |
| 650 | goto err; |
| 651 | |
| 652 | fio_cpuset_exit(&cpu_mask); |
| 653 | return NULL; |
| 654 | err: |
| 655 | fio_cpuset_exit(&cpu_mask); |
| 656 | err_out: |
| 657 | return (void *) 1; |
| 658 | } |
| 659 | |
| 660 | static int clock_cmp(const void *p1, const void *p2) |
| 661 | { |
| 662 | const struct clock_entry *c1 = p1; |
| 663 | const struct clock_entry *c2 = p2; |
| 664 | |
| 665 | if (c1->seq == c2->seq) |
| 666 | log_err("cs: bug in atomic sequence!\n"); |
| 667 | |
| 668 | return c1->seq - c2->seq; |
| 669 | } |
| 670 | |
| 671 | int fio_monotonic_clocktest(int debug) |
| 672 | { |
| 673 | struct clock_thread *cthreads; |
| 674 | unsigned int seen_cpus, nr_cpus = cpus_configured(); |
| 675 | struct clock_entry *entries; |
| 676 | unsigned long nr_entries, tentries, failed = 0; |
| 677 | struct clock_entry *prev, *this; |
| 678 | uint32_t seq = 0; |
| 679 | unsigned int i; |
| 680 | os_cpu_mask_t mask; |
| 681 | |
| 682 | #ifdef FIO_HAVE_GET_THREAD_AFFINITY |
| 683 | fio_get_thread_affinity(mask); |
| 684 | #else |
| 685 | memset(&mask, 0, sizeof(mask)); |
| 686 | for (i = 0; i < nr_cpus; i++) |
| 687 | fio_cpu_set(&mask, i); |
| 688 | #endif |
| 689 | |
| 690 | if (debug) { |
| 691 | log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no"); |
| 692 | |
| 693 | #ifdef FIO_INC_DEBUG |
| 694 | fio_debug |= 1U << FD_TIME; |
| 695 | #endif |
| 696 | nr_entries = CLOCK_ENTRIES_DEBUG; |
| 697 | } else |
| 698 | nr_entries = CLOCK_ENTRIES_TEST; |
| 699 | |
| 700 | calibrate_cpu_clock(); |
| 701 | |
| 702 | if (debug) { |
| 703 | #ifdef FIO_INC_DEBUG |
| 704 | fio_debug &= ~(1U << FD_TIME); |
| 705 | #endif |
| 706 | } |
| 707 | |
| 708 | cthreads = malloc(nr_cpus * sizeof(struct clock_thread)); |
| 709 | tentries = nr_entries * nr_cpus; |
| 710 | entries = malloc(tentries * sizeof(struct clock_entry)); |
| 711 | |
| 712 | if (debug) |
| 713 | log_info("cs: Testing %u CPUs\n", nr_cpus); |
| 714 | |
| 715 | seen_cpus = 0; |
| 716 | for (i = 0; i < nr_cpus; i++) { |
| 717 | struct clock_thread *t = &cthreads[i]; |
| 718 | |
| 719 | if (!fio_cpu_isset(&mask, i)) |
| 720 | continue; |
| 721 | t->cpu = i; |
| 722 | t->debug = debug; |
| 723 | t->seq = &seq; |
| 724 | t->nr_entries = nr_entries; |
| 725 | t->entries = &entries[seen_cpus * nr_entries]; |
| 726 | __fio_sem_init(&t->lock, FIO_SEM_LOCKED); |
| 727 | if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) { |
| 728 | failed++; |
| 729 | nr_cpus = i; |
| 730 | break; |
| 731 | } |
| 732 | seen_cpus++; |
| 733 | } |
| 734 | |
| 735 | for (i = 0; i < nr_cpus; i++) { |
| 736 | struct clock_thread *t = &cthreads[i]; |
| 737 | |
| 738 | if (!fio_cpu_isset(&mask, i)) |
| 739 | continue; |
| 740 | fio_sem_up(&t->lock); |
| 741 | } |
| 742 | |
| 743 | for (i = 0; i < nr_cpus; i++) { |
| 744 | struct clock_thread *t = &cthreads[i]; |
| 745 | void *ret; |
| 746 | |
| 747 | if (!fio_cpu_isset(&mask, i)) |
| 748 | continue; |
| 749 | pthread_join(t->thread, &ret); |
| 750 | if (ret) |
| 751 | failed++; |
| 752 | __fio_sem_remove(&t->lock); |
| 753 | } |
| 754 | free(cthreads); |
| 755 | |
| 756 | if (failed) { |
| 757 | if (debug) |
| 758 | log_err("Clocksource test: %lu threads failed\n", failed); |
| 759 | goto err; |
| 760 | } |
| 761 | |
| 762 | tentries = nr_entries * seen_cpus; |
| 763 | qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp); |
| 764 | |
| 765 | /* silence silly gcc */ |
| 766 | prev = NULL; |
| 767 | for (failed = i = 0; i < tentries; i++) { |
| 768 | this = &entries[i]; |
| 769 | |
| 770 | if (!i) { |
| 771 | prev = this; |
| 772 | continue; |
| 773 | } |
| 774 | |
| 775 | if (prev->tsc > this->tsc) { |
| 776 | uint64_t diff = prev->tsc - this->tsc; |
| 777 | |
| 778 | if (!debug) { |
| 779 | failed++; |
| 780 | break; |
| 781 | } |
| 782 | |
| 783 | log_info("cs: CPU clock mismatch (diff=%llu):\n", |
| 784 | (unsigned long long) diff); |
| 785 | log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq); |
| 786 | log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq); |
| 787 | failed++; |
| 788 | } |
| 789 | |
| 790 | prev = this; |
| 791 | } |
| 792 | |
| 793 | if (debug) { |
| 794 | if (failed) |
| 795 | log_info("cs: Failed: %lu\n", failed); |
| 796 | else |
| 797 | log_info("cs: Pass!\n"); |
| 798 | } |
| 799 | err: |
| 800 | free(entries); |
| 801 | return !!failed; |
| 802 | } |
| 803 | |
| 804 | #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */ |
| 805 | |
| 806 | int fio_monotonic_clocktest(int debug) |
| 807 | { |
| 808 | if (debug) |
| 809 | log_info("cs: current platform does not support CPU clocks\n"); |
| 810 | return 1; |
| 811 | } |
| 812 | |
| 813 | #endif |