1 #include <stdio.h> 2 #include <string.h> 3 #include <sys/time.h> 4 #include <sys/types.h> 5 #include <sys/stat.h> 6 #include <dirent.h> 7 #include <libgen.h> 8 #include <math.h> 9 10 #include "fio.h" 11 #include "diskutil.h" 12 #include "lib/ieee754.h" 13 #include "json.h" 14 #include "lib/getrusage.h" 15 #include "idletime.h" 16 17 static struct fio_mutex *stat_mutex; 18 19 void update_rusage_stat(struct thread_data *td) 20 { 21 struct thread_stat *ts = &td->ts; 22 23 fio_getrusage(&td->ru_end); 24 ts->usr_time += mtime_since(&td->ru_start.ru_utime, 25 &td->ru_end.ru_utime); 26 ts->sys_time += mtime_since(&td->ru_start.ru_stime, 27 &td->ru_end.ru_stime); 28 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw 29 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); 30 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt; 31 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt; 32 33 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); 34 } 35 36 /* 37 * Given a latency, return the index of the corresponding bucket in 38 * the structure tracking percentiles. 39 * 40 * (1) find the group (and error bits) that the value (latency) 41 * belongs to by looking at its MSB. (2) find the bucket number in the 42 * group by looking at the index bits. 43 * 44 */ 45 static unsigned int plat_val_to_idx(unsigned int val) 46 { 47 unsigned int msb, error_bits, base, offset, idx; 48 49 /* Find MSB starting from bit 0 */ 50 if (val == 0) 51 msb = 0; 52 else 53 msb = (sizeof(val)*8) - __builtin_clz(val) - 1; 54 55 /* 56 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 57 * all bits of the sample as index 58 */ 59 if (msb <= FIO_IO_U_PLAT_BITS) 60 return val; 61 62 /* Compute the number of error bits to discard*/ 63 error_bits = msb - FIO_IO_U_PLAT_BITS; 64 65 /* Compute the number of buckets before the group */ 66 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS; 67 68 /* 69 * Discard the error bits and apply the mask to find the 70 * index for the buckets in the group 71 */ 72 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits); 73 74 /* Make sure the index does not exceed (array size - 1) */ 75 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ? 76 (base + offset) : (FIO_IO_U_PLAT_NR - 1); 77 78 return idx; 79 } 80 81 /* 82 * Convert the given index of the bucket array to the value 83 * represented by the bucket 84 */ 85 static unsigned int plat_idx_to_val(unsigned int idx) 86 { 87 unsigned int error_bits, k, base; 88 89 assert(idx < FIO_IO_U_PLAT_NR); 90 91 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 92 * all bits of the sample as index */ 93 if (idx < (FIO_IO_U_PLAT_VAL << 1)) 94 return idx; 95 96 /* Find the group and compute the minimum value of that group */ 97 error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1; 98 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS); 99 100 /* Find its bucket number of the group */ 101 k = idx % FIO_IO_U_PLAT_VAL; 102 103 /* Return the mean of the range of the bucket */ 104 return base + ((k + 0.5) * (1 << error_bits)); 105 } 106 107 static int double_cmp(const void *a, const void *b) 108 { 109 const fio_fp64_t fa = *(const fio_fp64_t *) a; 110 const fio_fp64_t fb = *(const fio_fp64_t *) b; 111 int cmp = 0; 112 113 if (fa.u.f > fb.u.f) 114 cmp = 1; 115 else if (fa.u.f < fb.u.f) 116 cmp = -1; 117 118 return cmp; 119 } 120 121 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 122 fio_fp64_t *plist, unsigned int **output, 123 unsigned int *maxv, unsigned int *minv) 124 { 125 unsigned long sum = 0; 126 unsigned int len, i, j = 0; 127 unsigned int oval_len = 0; 128 unsigned int *ovals = NULL; 129 int is_last; 130 131 *minv = -1U; 132 *maxv = 0; 133 134 len = 0; 135 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) 136 len++; 137 138 if (!len) 139 return 0; 140 141 /* 142 * Sort the percentile list. Note that it may already be sorted if 143 * we are using the default values, but since it's a short list this 144 * isn't a worry. Also note that this does not work for NaN values. 145 */ 146 if (len > 1) 147 qsort((void *)plist, len, sizeof(plist[0]), double_cmp); 148 149 /* 150 * Calculate bucket values, note down max and min values 151 */ 152 is_last = 0; 153 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) { 154 sum += io_u_plat[i]; 155 while (sum >= (plist[j].u.f / 100.0 * nr)) { 156 assert(plist[j].u.f <= 100.0); 157 158 if (j == oval_len) { 159 oval_len += 100; 160 ovals = realloc(ovals, oval_len * sizeof(unsigned int)); 161 } 162 163 ovals[j] = plat_idx_to_val(i); 164 if (ovals[j] < *minv) 165 *minv = ovals[j]; 166 if (ovals[j] > *maxv) 167 *maxv = ovals[j]; 168 169 is_last = (j == len - 1); 170 if (is_last) 171 break; 172 173 j++; 174 } 175 } 176 177 *output = ovals; 178 return len; 179 } 180 181 /* 182 * Find and display the p-th percentile of clat 183 */ 184 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 185 fio_fp64_t *plist, unsigned int precision) 186 { 187 unsigned int len, j = 0, minv, maxv; 188 unsigned int *ovals; 189 int is_last, per_line, scale_down; 190 char fmt[32]; 191 192 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv); 193 if (!len) 194 goto out; 195 196 /* 197 * We default to usecs, but if the value range is such that we 198 * should scale down to msecs, do that. 199 */ 200 if (minv > 2000 && maxv > 99999) { 201 scale_down = 1; 202 log_info(" clat percentiles (msec):\n |"); 203 } else { 204 scale_down = 0; 205 log_info(" clat percentiles (usec):\n |"); 206 } 207 208 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision); 209 per_line = (80 - 7) / (precision + 14); 210 211 for (j = 0; j < len; j++) { 212 char fbuf[16], *ptr = fbuf; 213 214 /* for formatting */ 215 if (j != 0 && (j % per_line) == 0) 216 log_info(" |"); 217 218 /* end of the list */ 219 is_last = (j == len - 1); 220 221 if (plist[j].u.f < 10.0) 222 ptr += sprintf(fbuf, " "); 223 224 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f); 225 226 if (scale_down) 227 ovals[j] = (ovals[j] + 999) / 1000; 228 229 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ','); 230 231 if (is_last) 232 break; 233 234 if ((j % per_line) == per_line - 1) /* for formatting */ 235 log_info("\n"); 236 } 237 238 out: 239 if (ovals) 240 free(ovals); 241 } 242 243 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, 244 double *mean, double *dev) 245 { 246 double n = (double) is->samples; 247 248 if (n == 0) 249 return 0; 250 251 *min = is->min_val; 252 *max = is->max_val; 253 *mean = is->mean.u.f; 254 255 if (n > 1.0) 256 *dev = sqrt(is->S.u.f / (n - 1.0)); 257 else 258 *dev = 0; 259 260 return 1; 261 } 262 263 void show_group_stats(struct group_run_stats *rs) 264 { 265 char *p1, *p2, *p3, *p4; 266 const char *ddir_str[] = { " READ", " WRITE" , " TRIM"}; 267 int i; 268 269 log_info("\nRun status group %d (all jobs):\n", rs->groupid); 270 271 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 272 const int i2p = is_power_of_2(rs->kb_base); 273 274 if (!rs->max_run[i]) 275 continue; 276 277 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8); 278 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base); 279 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base); 280 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base); 281 282 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s," 283 " mint=%llumsec, maxt=%llumsec\n", 284 rs->unified_rw_rep ? " MIXED" : ddir_str[i], 285 p1, p2, p3, p4, 286 (unsigned long long) rs->min_run[i], 287 (unsigned long long) rs->max_run[i]); 288 289 free(p1); 290 free(p2); 291 free(p3); 292 free(p4); 293 } 294 } 295 296 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist) 297 { 298 int i; 299 300 /* 301 * Do depth distribution calculations 302 */ 303 for (i = 0; i < FIO_IO_U_MAP_NR; i++) { 304 if (total) { 305 io_u_dist[i] = (double) map[i] / (double) total; 306 io_u_dist[i] *= 100.0; 307 if (io_u_dist[i] < 0.1 && map[i]) 308 io_u_dist[i] = 0.1; 309 } else 310 io_u_dist[i] = 0.0; 311 } 312 } 313 314 static void stat_calc_lat(struct thread_stat *ts, double *dst, 315 unsigned int *src, int nr) 316 { 317 unsigned long total = ddir_rw_sum(ts->total_io_u); 318 int i; 319 320 /* 321 * Do latency distribution calculations 322 */ 323 for (i = 0; i < nr; i++) { 324 if (total) { 325 dst[i] = (double) src[i] / (double) total; 326 dst[i] *= 100.0; 327 if (dst[i] < 0.01 && src[i]) 328 dst[i] = 0.01; 329 } else 330 dst[i] = 0.0; 331 } 332 } 333 334 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat) 335 { 336 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR); 337 } 338 339 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat) 340 { 341 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR); 342 } 343 344 static void display_lat(const char *name, unsigned long min, unsigned long max, 345 double mean, double dev) 346 { 347 const char *base = "(usec)"; 348 char *minp, *maxp; 349 350 if (!usec_to_msec(&min, &max, &mean, &dev)) 351 base = "(msec)"; 352 353 minp = num2str(min, 6, 1, 0, 0); 354 maxp = num2str(max, 6, 1, 0, 0); 355 356 log_info(" %s %s: min=%s, max=%s, avg=%5.02f," 357 " stdev=%5.02f\n", name, base, minp, maxp, mean, dev); 358 359 free(minp); 360 free(maxp); 361 } 362 363 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts, 364 int ddir) 365 { 366 const char *ddir_str[] = { "read ", "write", "trim" }; 367 unsigned long min, max, runt; 368 unsigned long long bw, iops; 369 double mean, dev; 370 char *io_p, *bw_p, *iops_p; 371 int i2p; 372 373 assert(ddir_rw(ddir)); 374 375 if (!ts->runtime[ddir]) 376 return; 377 378 i2p = is_power_of_2(rs->kb_base); 379 runt = ts->runtime[ddir]; 380 381 bw = (1000 * ts->io_bytes[ddir]) / runt; 382 io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8); 383 bw_p = num2str(bw, 6, 1, i2p, ts->unit_base); 384 385 iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt; 386 iops_p = num2str(iops, 6, 1, 0, 0); 387 388 log_info(" %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n", 389 rs->unified_rw_rep ? "mixed" : ddir_str[ddir], 390 io_p, bw_p, iops_p, 391 (unsigned long long) ts->runtime[ddir]); 392 393 free(io_p); 394 free(bw_p); 395 free(iops_p); 396 397 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 398 display_lat("slat", min, max, mean, dev); 399 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 400 display_lat("clat", min, max, mean, dev); 401 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 402 display_lat(" lat", min, max, mean, dev); 403 404 if (ts->clat_percentiles) { 405 show_clat_percentiles(ts->io_u_plat[ddir], 406 ts->clat_stat[ddir].samples, 407 ts->percentile_list, 408 ts->percentile_precision); 409 } 410 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 411 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base; 412 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB"); 413 414 if (rs->unit_base == 1) { 415 min *= 8.0; 416 max *= 8.0; 417 mean *= 8.0; 418 dev *= 8.0; 419 } 420 421 if (rs->agg[ddir]) { 422 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 423 if (p_of_agg > 100.0) 424 p_of_agg = 100.0; 425 } 426 427 if (mean > fkb_base * fkb_base) { 428 min /= fkb_base; 429 max /= fkb_base; 430 mean /= fkb_base; 431 dev /= fkb_base; 432 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB"); 433 } 434 435 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%," 436 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max, 437 p_of_agg, mean, dev); 438 } 439 } 440 441 static int show_lat(double *io_u_lat, int nr, const char **ranges, 442 const char *msg) 443 { 444 int new_line = 1, i, line = 0, shown = 0; 445 446 for (i = 0; i < nr; i++) { 447 if (io_u_lat[i] <= 0.0) 448 continue; 449 shown = 1; 450 if (new_line) { 451 if (line) 452 log_info("\n"); 453 log_info(" lat (%s) : ", msg); 454 new_line = 0; 455 line = 0; 456 } 457 if (line) 458 log_info(", "); 459 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]); 460 line++; 461 if (line == 5) 462 new_line = 1; 463 } 464 465 if (shown) 466 log_info("\n"); 467 468 return shown; 469 } 470 471 static void show_lat_u(double *io_u_lat_u) 472 { 473 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 474 "250=", "500=", "750=", "1000=", }; 475 476 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec"); 477 } 478 479 static void show_lat_m(double *io_u_lat_m) 480 { 481 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 482 "250=", "500=", "750=", "1000=", "2000=", 483 ">=2000=", }; 484 485 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec"); 486 } 487 488 static void show_latencies(struct thread_stat *ts) 489 { 490 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 491 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 492 493 stat_calc_lat_u(ts, io_u_lat_u); 494 stat_calc_lat_m(ts, io_u_lat_m); 495 496 show_lat_u(io_u_lat_u); 497 show_lat_m(io_u_lat_m); 498 } 499 500 static void show_thread_status_normal(struct thread_stat *ts, 501 struct group_run_stats *rs) 502 { 503 double usr_cpu, sys_cpu; 504 unsigned long runtime; 505 double io_u_dist[FIO_IO_U_MAP_NR]; 506 time_t time_p; 507 char time_buf[64]; 508 509 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] + 510 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] + 511 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM])) 512 return; 513 514 time(&time_p); 515 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf)); 516 517 if (!ts->error) { 518 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s", 519 ts->name, ts->groupid, ts->members, 520 ts->error, (int) ts->pid, time_buf); 521 } else { 522 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s", 523 ts->name, ts->groupid, ts->members, 524 ts->error, ts->verror, (int) ts->pid, 525 time_buf); 526 } 527 528 if (strlen(ts->description)) 529 log_info(" Description : [%s]\n", ts->description); 530 531 if (ts->io_bytes[DDIR_READ]) 532 show_ddir_status(rs, ts, DDIR_READ); 533 if (ts->io_bytes[DDIR_WRITE]) 534 show_ddir_status(rs, ts, DDIR_WRITE); 535 if (ts->io_bytes[DDIR_TRIM]) 536 show_ddir_status(rs, ts, DDIR_TRIM); 537 538 show_latencies(ts); 539 540 runtime = ts->total_run_time; 541 if (runtime) { 542 double runt = (double) runtime; 543 544 usr_cpu = (double) ts->usr_time * 100 / runt; 545 sys_cpu = (double) ts->sys_time * 100 / runt; 546 } else { 547 usr_cpu = 0; 548 sys_cpu = 0; 549 } 550 551 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu," 552 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu, 553 (unsigned long long) ts->ctx, 554 (unsigned long long) ts->majf, 555 (unsigned long long) ts->minf); 556 557 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 558 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 559 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 560 io_u_dist[1], io_u_dist[2], 561 io_u_dist[3], io_u_dist[4], 562 io_u_dist[5], io_u_dist[6]); 563 564 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 565 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 566 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 567 io_u_dist[1], io_u_dist[2], 568 io_u_dist[3], io_u_dist[4], 569 io_u_dist[5], io_u_dist[6]); 570 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 571 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 572 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 573 io_u_dist[1], io_u_dist[2], 574 io_u_dist[3], io_u_dist[4], 575 io_u_dist[5], io_u_dist[6]); 576 log_info(" issued : total=r=%llu/w=%llu/d=%llu," 577 " short=r=%llu/w=%llu/d=%llu\n", 578 (unsigned long long) ts->total_io_u[0], 579 (unsigned long long) ts->total_io_u[1], 580 (unsigned long long) ts->total_io_u[2], 581 (unsigned long long) ts->short_io_u[0], 582 (unsigned long long) ts->short_io_u[1], 583 (unsigned long long) ts->short_io_u[2]); 584 if (ts->continue_on_error) { 585 log_info(" errors : total=%llu, first_error=%d/<%s>\n", 586 (unsigned long long)ts->total_err_count, 587 ts->first_error, 588 strerror(ts->first_error)); 589 } 590 if (ts->latency_depth) { 591 log_info(" latency : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n", 592 (unsigned long long)ts->latency_target, 593 (unsigned long long)ts->latency_window, 594 ts->latency_percentile.u.f, 595 ts->latency_depth); 596 } 597 } 598 599 static void show_ddir_status_terse(struct thread_stat *ts, 600 struct group_run_stats *rs, int ddir) 601 { 602 unsigned long min, max; 603 unsigned long long bw, iops; 604 unsigned int *ovals = NULL; 605 double mean, dev; 606 unsigned int len, minv, maxv; 607 int i; 608 609 assert(ddir_rw(ddir)); 610 611 iops = bw = 0; 612 if (ts->runtime[ddir]) { 613 uint64_t runt = ts->runtime[ddir]; 614 615 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 616 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 617 } 618 619 log_info(";%llu;%llu;%llu;%llu", 620 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops, 621 (unsigned long long) ts->runtime[ddir]); 622 623 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 624 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 625 else 626 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 627 628 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 629 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 630 else 631 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 632 633 if (ts->clat_percentiles) { 634 len = calc_clat_percentiles(ts->io_u_plat[ddir], 635 ts->clat_stat[ddir].samples, 636 ts->percentile_list, &ovals, &maxv, 637 &minv); 638 } else 639 len = 0; 640 641 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 642 if (i >= len) { 643 log_info(";0%%=0"); 644 continue; 645 } 646 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]); 647 } 648 649 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 650 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 651 else 652 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 653 654 if (ovals) 655 free(ovals); 656 657 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 658 double p_of_agg = 100.0; 659 660 if (rs->agg[ddir]) { 661 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 662 if (p_of_agg > 100.0) 663 p_of_agg = 100.0; 664 } 665 666 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 667 } else 668 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 669 } 670 671 static void add_ddir_status_json(struct thread_stat *ts, 672 struct group_run_stats *rs, int ddir, struct json_object *parent) 673 { 674 unsigned long min, max; 675 unsigned long long bw, iops; 676 unsigned int *ovals = NULL; 677 double mean, dev; 678 unsigned int len, minv, maxv; 679 int i; 680 const char *ddirname[] = {"read", "write", "trim"}; 681 struct json_object *dir_object, *tmp_object, *percentile_object; 682 char buf[120]; 683 double p_of_agg = 100.0; 684 685 assert(ddir_rw(ddir)); 686 687 if (ts->unified_rw_rep && ddir != DDIR_READ) 688 return; 689 690 dir_object = json_create_object(); 691 json_object_add_value_object(parent, 692 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object); 693 694 iops = bw = 0; 695 if (ts->runtime[ddir]) { 696 uint64_t runt = ts->runtime[ddir]; 697 698 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 699 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 700 } 701 702 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10); 703 json_object_add_value_int(dir_object, "bw", bw); 704 json_object_add_value_int(dir_object, "iops", iops); 705 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]); 706 707 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 708 min = max = 0; 709 mean = dev = 0.0; 710 } 711 tmp_object = json_create_object(); 712 json_object_add_value_object(dir_object, "slat", tmp_object); 713 json_object_add_value_int(tmp_object, "min", min); 714 json_object_add_value_int(tmp_object, "max", max); 715 json_object_add_value_float(tmp_object, "mean", mean); 716 json_object_add_value_float(tmp_object, "stddev", dev); 717 718 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 719 min = max = 0; 720 mean = dev = 0.0; 721 } 722 tmp_object = json_create_object(); 723 json_object_add_value_object(dir_object, "clat", tmp_object); 724 json_object_add_value_int(tmp_object, "min", min); 725 json_object_add_value_int(tmp_object, "max", max); 726 json_object_add_value_float(tmp_object, "mean", mean); 727 json_object_add_value_float(tmp_object, "stddev", dev); 728 729 if (ts->clat_percentiles) { 730 len = calc_clat_percentiles(ts->io_u_plat[ddir], 731 ts->clat_stat[ddir].samples, 732 ts->percentile_list, &ovals, &maxv, 733 &minv); 734 } else 735 len = 0; 736 737 percentile_object = json_create_object(); 738 json_object_add_value_object(tmp_object, "percentile", percentile_object); 739 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 740 if (i >= len) { 741 json_object_add_value_int(percentile_object, "0.00", 0); 742 continue; 743 } 744 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f); 745 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]); 746 } 747 748 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 749 min = max = 0; 750 mean = dev = 0.0; 751 } 752 tmp_object = json_create_object(); 753 json_object_add_value_object(dir_object, "lat", tmp_object); 754 json_object_add_value_int(tmp_object, "min", min); 755 json_object_add_value_int(tmp_object, "max", max); 756 json_object_add_value_float(tmp_object, "mean", mean); 757 json_object_add_value_float(tmp_object, "stddev", dev); 758 if (ovals) 759 free(ovals); 760 761 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 762 if (rs->agg[ddir]) { 763 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 764 if (p_of_agg > 100.0) 765 p_of_agg = 100.0; 766 } 767 } else { 768 min = max = 0; 769 p_of_agg = mean = dev = 0.0; 770 } 771 json_object_add_value_int(dir_object, "bw_min", min); 772 json_object_add_value_int(dir_object, "bw_max", max); 773 json_object_add_value_float(dir_object, "bw_agg", p_of_agg); 774 json_object_add_value_float(dir_object, "bw_mean", mean); 775 json_object_add_value_float(dir_object, "bw_dev", dev); 776 } 777 778 static void show_thread_status_terse_v2(struct thread_stat *ts, 779 struct group_run_stats *rs) 780 { 781 double io_u_dist[FIO_IO_U_MAP_NR]; 782 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 783 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 784 double usr_cpu, sys_cpu; 785 int i; 786 787 /* General Info */ 788 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error); 789 /* Log Read Status */ 790 show_ddir_status_terse(ts, rs, DDIR_READ); 791 /* Log Write Status */ 792 show_ddir_status_terse(ts, rs, DDIR_WRITE); 793 /* Log Trim Status */ 794 show_ddir_status_terse(ts, rs, DDIR_TRIM); 795 796 /* CPU Usage */ 797 if (ts->total_run_time) { 798 double runt = (double) ts->total_run_time; 799 800 usr_cpu = (double) ts->usr_time * 100 / runt; 801 sys_cpu = (double) ts->sys_time * 100 / runt; 802 } else { 803 usr_cpu = 0; 804 sys_cpu = 0; 805 } 806 807 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, 808 (unsigned long long) ts->ctx, 809 (unsigned long long) ts->majf, 810 (unsigned long long) ts->minf); 811 812 /* Calc % distribution of IO depths, usecond, msecond latency */ 813 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 814 stat_calc_lat_u(ts, io_u_lat_u); 815 stat_calc_lat_m(ts, io_u_lat_m); 816 817 /* Only show fixed 7 I/O depth levels*/ 818 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 819 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 820 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 821 822 /* Microsecond latency */ 823 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 824 log_info(";%3.2f%%", io_u_lat_u[i]); 825 /* Millisecond latency */ 826 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 827 log_info(";%3.2f%%", io_u_lat_m[i]); 828 /* Additional output if continue_on_error set - default off*/ 829 if (ts->continue_on_error) 830 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); 831 log_info("\n"); 832 833 /* Additional output if description is set */ 834 if (strlen(ts->description)) 835 log_info(";%s", ts->description); 836 837 log_info("\n"); 838 } 839 840 static void show_thread_status_terse_v3_v4(struct thread_stat *ts, 841 struct group_run_stats *rs, int ver) 842 { 843 double io_u_dist[FIO_IO_U_MAP_NR]; 844 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 845 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 846 double usr_cpu, sys_cpu; 847 int i; 848 849 /* General Info */ 850 log_info("%d;%s;%s;%d;%d", ver, fio_version_string, 851 ts->name, ts->groupid, ts->error); 852 /* Log Read Status */ 853 show_ddir_status_terse(ts, rs, DDIR_READ); 854 /* Log Write Status */ 855 show_ddir_status_terse(ts, rs, DDIR_WRITE); 856 /* Log Trim Status */ 857 if (ver == 4) 858 show_ddir_status_terse(ts, rs, DDIR_TRIM); 859 860 /* CPU Usage */ 861 if (ts->total_run_time) { 862 double runt = (double) ts->total_run_time; 863 864 usr_cpu = (double) ts->usr_time * 100 / runt; 865 sys_cpu = (double) ts->sys_time * 100 / runt; 866 } else { 867 usr_cpu = 0; 868 sys_cpu = 0; 869 } 870 871 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, 872 (unsigned long long) ts->ctx, 873 (unsigned long long) ts->majf, 874 (unsigned long long) ts->minf); 875 876 /* Calc % distribution of IO depths, usecond, msecond latency */ 877 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 878 stat_calc_lat_u(ts, io_u_lat_u); 879 stat_calc_lat_m(ts, io_u_lat_m); 880 881 /* Only show fixed 7 I/O depth levels*/ 882 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 883 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 884 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 885 886 /* Microsecond latency */ 887 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 888 log_info(";%3.2f%%", io_u_lat_u[i]); 889 /* Millisecond latency */ 890 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 891 log_info(";%3.2f%%", io_u_lat_m[i]); 892 893 /* disk util stats, if any */ 894 if (is_backend) 895 show_disk_util(1, NULL); 896 897 /* Additional output if continue_on_error set - default off*/ 898 if (ts->continue_on_error) 899 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); 900 901 /* Additional output if description is set */ 902 if (strlen(ts->description)) 903 log_info(";%s", ts->description); 904 905 log_info("\n"); 906 } 907 908 static struct json_object *show_thread_status_json(struct thread_stat *ts, 909 struct group_run_stats *rs) 910 { 911 struct json_object *root, *tmp; 912 double io_u_dist[FIO_IO_U_MAP_NR]; 913 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 914 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 915 double usr_cpu, sys_cpu; 916 int i; 917 918 root = json_create_object(); 919 json_object_add_value_string(root, "jobname", ts->name); 920 json_object_add_value_int(root, "groupid", ts->groupid); 921 json_object_add_value_int(root, "error", ts->error); 922 923 add_ddir_status_json(ts, rs, DDIR_READ, root); 924 add_ddir_status_json(ts, rs, DDIR_WRITE, root); 925 add_ddir_status_json(ts, rs, DDIR_TRIM, root); 926 927 /* CPU Usage */ 928 if (ts->total_run_time) { 929 double runt = (double) ts->total_run_time; 930 931 usr_cpu = (double) ts->usr_time * 100 / runt; 932 sys_cpu = (double) ts->sys_time * 100 / runt; 933 } else { 934 usr_cpu = 0; 935 sys_cpu = 0; 936 } 937 json_object_add_value_float(root, "usr_cpu", usr_cpu); 938 json_object_add_value_float(root, "sys_cpu", sys_cpu); 939 json_object_add_value_int(root, "ctx", ts->ctx); 940 json_object_add_value_int(root, "majf", ts->majf); 941 json_object_add_value_int(root, "minf", ts->minf); 942 943 944 /* Calc % distribution of IO depths, usecond, msecond latency */ 945 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 946 stat_calc_lat_u(ts, io_u_lat_u); 947 stat_calc_lat_m(ts, io_u_lat_m); 948 949 tmp = json_create_object(); 950 json_object_add_value_object(root, "iodepth_level", tmp); 951 /* Only show fixed 7 I/O depth levels*/ 952 for (i = 0; i < 7; i++) { 953 char name[20]; 954 if (i < 6) 955 snprintf(name, 20, "%d", 1 << i); 956 else 957 snprintf(name, 20, ">=%d", 1 << i); 958 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]); 959 } 960 961 tmp = json_create_object(); 962 json_object_add_value_object(root, "latency_us", tmp); 963 /* Microsecond latency */ 964 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) { 965 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 966 "250", "500", "750", "1000", }; 967 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]); 968 } 969 /* Millisecond latency */ 970 tmp = json_create_object(); 971 json_object_add_value_object(root, "latency_ms", tmp); 972 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) { 973 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 974 "250", "500", "750", "1000", "2000", 975 ">=2000", }; 976 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]); 977 } 978 979 /* Additional output if continue_on_error set - default off*/ 980 if (ts->continue_on_error) { 981 json_object_add_value_int(root, "total_err", ts->total_err_count); 982 json_object_add_value_int(root, "first_error", ts->first_error); 983 } 984 985 if (ts->latency_depth) { 986 json_object_add_value_int(root, "latency_depth", ts->latency_depth); 987 json_object_add_value_int(root, "latency_target", ts->latency_target); 988 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f); 989 json_object_add_value_int(root, "latency_window", ts->latency_window); 990 } 991 992 /* Additional output if description is set */ 993 if (strlen(ts->description)) 994 json_object_add_value_string(root, "desc", ts->description); 995 996 return root; 997 } 998 999 static void show_thread_status_terse(struct thread_stat *ts, 1000 struct group_run_stats *rs) 1001 { 1002 if (terse_version == 2) 1003 show_thread_status_terse_v2(ts, rs); 1004 else if (terse_version == 3 || terse_version == 4) 1005 show_thread_status_terse_v3_v4(ts, rs, terse_version); 1006 else 1007 log_err("fio: bad terse version!? %d\n", terse_version); 1008 } 1009 1010 struct json_object *show_thread_status(struct thread_stat *ts, 1011 struct group_run_stats *rs) 1012 { 1013 if (output_format == FIO_OUTPUT_TERSE) 1014 show_thread_status_terse(ts, rs); 1015 else if (output_format == FIO_OUTPUT_JSON) 1016 return show_thread_status_json(ts, rs); 1017 else 1018 show_thread_status_normal(ts, rs); 1019 return NULL; 1020 } 1021 1022 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr) 1023 { 1024 double mean, S; 1025 1026 if (src->samples == 0) 1027 return; 1028 1029 dst->min_val = min(dst->min_val, src->min_val); 1030 dst->max_val = max(dst->max_val, src->max_val); 1031 1032 /* 1033 * Compute new mean and S after the merge 1034 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 1035 * #Parallel_algorithm> 1036 */ 1037 if (nr == 1) { 1038 mean = src->mean.u.f; 1039 S = src->S.u.f; 1040 } else { 1041 double delta = src->mean.u.f - dst->mean.u.f; 1042 1043 mean = ((src->mean.u.f * src->samples) + 1044 (dst->mean.u.f * dst->samples)) / 1045 (dst->samples + src->samples); 1046 1047 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * 1048 (dst->samples * src->samples) / 1049 (dst->samples + src->samples); 1050 } 1051 1052 dst->samples += src->samples; 1053 dst->mean.u.f = mean; 1054 dst->S.u.f = S; 1055 } 1056 1057 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) 1058 { 1059 int i; 1060 1061 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 1062 if (dst->max_run[i] < src->max_run[i]) 1063 dst->max_run[i] = src->max_run[i]; 1064 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) 1065 dst->min_run[i] = src->min_run[i]; 1066 if (dst->max_bw[i] < src->max_bw[i]) 1067 dst->max_bw[i] = src->max_bw[i]; 1068 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) 1069 dst->min_bw[i] = src->min_bw[i]; 1070 1071 dst->io_kb[i] += src->io_kb[i]; 1072 dst->agg[i] += src->agg[i]; 1073 } 1074 1075 } 1076 1077 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr) 1078 { 1079 int l, k; 1080 1081 for (l = 0; l < DDIR_RWDIR_CNT; l++) { 1082 if (!dst->unified_rw_rep) { 1083 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr); 1084 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr); 1085 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr); 1086 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr); 1087 1088 dst->io_bytes[l] += src->io_bytes[l]; 1089 1090 if (dst->runtime[l] < src->runtime[l]) 1091 dst->runtime[l] = src->runtime[l]; 1092 } else { 1093 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr); 1094 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr); 1095 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr); 1096 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr); 1097 1098 dst->io_bytes[0] += src->io_bytes[l]; 1099 1100 if (dst->runtime[0] < src->runtime[l]) 1101 dst->runtime[0] = src->runtime[l]; 1102 } 1103 } 1104 1105 dst->usr_time += src->usr_time; 1106 dst->sys_time += src->sys_time; 1107 dst->ctx += src->ctx; 1108 dst->majf += src->majf; 1109 dst->minf += src->minf; 1110 1111 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1112 dst->io_u_map[k] += src->io_u_map[k]; 1113 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1114 dst->io_u_submit[k] += src->io_u_submit[k]; 1115 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1116 dst->io_u_complete[k] += src->io_u_complete[k]; 1117 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) 1118 dst->io_u_lat_u[k] += src->io_u_lat_u[k]; 1119 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) 1120 dst->io_u_lat_m[k] += src->io_u_lat_m[k]; 1121 1122 for (k = 0; k < DDIR_RWDIR_CNT; k++) { 1123 if (!dst->unified_rw_rep) { 1124 dst->total_io_u[k] += src->total_io_u[k]; 1125 dst->short_io_u[k] += src->short_io_u[k]; 1126 } else { 1127 dst->total_io_u[0] += src->total_io_u[k]; 1128 dst->short_io_u[0] += src->short_io_u[k]; 1129 } 1130 } 1131 1132 for (k = 0; k < DDIR_RWDIR_CNT; k++) { 1133 int m; 1134 1135 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) { 1136 if (!dst->unified_rw_rep) 1137 dst->io_u_plat[k][m] += src->io_u_plat[k][m]; 1138 else 1139 dst->io_u_plat[0][m] += src->io_u_plat[k][m]; 1140 } 1141 } 1142 1143 dst->total_run_time += src->total_run_time; 1144 dst->total_submit += src->total_submit; 1145 dst->total_complete += src->total_complete; 1146 } 1147 1148 void init_group_run_stat(struct group_run_stats *gs) 1149 { 1150 int i; 1151 memset(gs, 0, sizeof(*gs)); 1152 1153 for (i = 0; i < DDIR_RWDIR_CNT; i++) 1154 gs->min_bw[i] = gs->min_run[i] = ~0UL; 1155 } 1156 1157 void init_thread_stat(struct thread_stat *ts) 1158 { 1159 int j; 1160 1161 memset(ts, 0, sizeof(*ts)); 1162 1163 for (j = 0; j < DDIR_RWDIR_CNT; j++) { 1164 ts->lat_stat[j].min_val = -1UL; 1165 ts->clat_stat[j].min_val = -1UL; 1166 ts->slat_stat[j].min_val = -1UL; 1167 ts->bw_stat[j].min_val = -1UL; 1168 } 1169 ts->groupid = -1; 1170 } 1171 1172 static void __show_run_stats(void) 1173 { 1174 struct group_run_stats *runstats, *rs; 1175 struct thread_data *td; 1176 struct thread_stat *threadstats, *ts; 1177 int i, j, nr_ts, last_ts, idx; 1178 int kb_base_warned = 0; 1179 int unit_base_warned = 0; 1180 struct json_object *root = NULL; 1181 struct json_array *array = NULL; 1182 1183 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); 1184 1185 for (i = 0; i < groupid + 1; i++) 1186 init_group_run_stat(&runstats[i]); 1187 1188 /* 1189 * find out how many threads stats we need. if group reporting isn't 1190 * enabled, it's one-per-td. 1191 */ 1192 nr_ts = 0; 1193 last_ts = -1; 1194 for_each_td(td, i) { 1195 if (!td->o.group_reporting) { 1196 nr_ts++; 1197 continue; 1198 } 1199 if (last_ts == td->groupid) 1200 continue; 1201 1202 last_ts = td->groupid; 1203 nr_ts++; 1204 } 1205 1206 threadstats = malloc(nr_ts * sizeof(struct thread_stat)); 1207 1208 for (i = 0; i < nr_ts; i++) 1209 init_thread_stat(&threadstats[i]); 1210 1211 j = 0; 1212 last_ts = -1; 1213 idx = 0; 1214 for_each_td(td, i) { 1215 if (idx && (!td->o.group_reporting || 1216 (td->o.group_reporting && last_ts != td->groupid))) { 1217 idx = 0; 1218 j++; 1219 } 1220 1221 last_ts = td->groupid; 1222 1223 ts = &threadstats[j]; 1224 1225 ts->clat_percentiles = td->o.clat_percentiles; 1226 ts->percentile_precision = td->o.percentile_precision; 1227 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); 1228 1229 idx++; 1230 ts->members++; 1231 1232 if (ts->groupid == -1) { 1233 /* 1234 * These are per-group shared already 1235 */ 1236 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1); 1237 if (td->o.description) 1238 strncpy(ts->description, td->o.description, 1239 FIO_JOBDESC_SIZE - 1); 1240 else 1241 memset(ts->description, 0, FIO_JOBDESC_SIZE); 1242 1243 /* 1244 * If multiple entries in this group, this is 1245 * the first member. 1246 */ 1247 ts->thread_number = td->thread_number; 1248 ts->groupid = td->groupid; 1249 1250 /* 1251 * first pid in group, not very useful... 1252 */ 1253 ts->pid = td->pid; 1254 1255 ts->kb_base = td->o.kb_base; 1256 ts->unit_base = td->o.unit_base; 1257 ts->unified_rw_rep = td->o.unified_rw_rep; 1258 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 1259 log_info("fio: kb_base differs for jobs in group, using" 1260 " %u as the base\n", ts->kb_base); 1261 kb_base_warned = 1; 1262 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) { 1263 log_info("fio: unit_base differs for jobs in group, using" 1264 " %u as the base\n", ts->unit_base); 1265 unit_base_warned = 1; 1266 } 1267 1268 ts->continue_on_error = td->o.continue_on_error; 1269 ts->total_err_count += td->total_err_count; 1270 ts->first_error = td->first_error; 1271 if (!ts->error) { 1272 if (!td->error && td->o.continue_on_error && 1273 td->first_error) { 1274 ts->error = td->first_error; 1275 ts->verror[sizeof(ts->verror) - 1] = '\0'; 1276 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1); 1277 } else if (td->error) { 1278 ts->error = td->error; 1279 ts->verror[sizeof(ts->verror) - 1] = '\0'; 1280 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1); 1281 } 1282 } 1283 1284 ts->latency_depth = td->latency_qd; 1285 ts->latency_target = td->o.latency_target; 1286 ts->latency_percentile = td->o.latency_percentile; 1287 ts->latency_window = td->o.latency_window; 1288 1289 sum_thread_stats(ts, &td->ts, idx); 1290 } 1291 1292 for (i = 0; i < nr_ts; i++) { 1293 unsigned long long bw; 1294 1295 ts = &threadstats[i]; 1296 rs = &runstats[ts->groupid]; 1297 rs->kb_base = ts->kb_base; 1298 rs->unit_base = ts->unit_base; 1299 rs->unified_rw_rep += ts->unified_rw_rep; 1300 1301 for (j = 0; j < DDIR_RWDIR_CNT; j++) { 1302 if (!ts->runtime[j]) 1303 continue; 1304 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 1305 rs->min_run[j] = ts->runtime[j]; 1306 if (ts->runtime[j] > rs->max_run[j]) 1307 rs->max_run[j] = ts->runtime[j]; 1308 1309 bw = 0; 1310 if (ts->runtime[j]) { 1311 unsigned long runt = ts->runtime[j]; 1312 unsigned long long kb; 1313 1314 kb = ts->io_bytes[j] / rs->kb_base; 1315 bw = kb * 1000 / runt; 1316 } 1317 if (bw < rs->min_bw[j]) 1318 rs->min_bw[j] = bw; 1319 if (bw > rs->max_bw[j]) 1320 rs->max_bw[j] = bw; 1321 1322 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 1323 } 1324 } 1325 1326 for (i = 0; i < groupid + 1; i++) { 1327 int ddir; 1328 1329 rs = &runstats[i]; 1330 1331 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) { 1332 if (rs->max_run[ddir]) 1333 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) / 1334 rs->max_run[ddir]; 1335 } 1336 } 1337 1338 /* 1339 * don't overwrite last signal output 1340 */ 1341 if (output_format == FIO_OUTPUT_NORMAL) 1342 log_info("\n"); 1343 else if (output_format == FIO_OUTPUT_JSON) { 1344 root = json_create_object(); 1345 json_object_add_value_string(root, "fio version", fio_version_string); 1346 array = json_create_array(); 1347 json_object_add_value_array(root, "jobs", array); 1348 } 1349 1350 for (i = 0; i < nr_ts; i++) { 1351 ts = &threadstats[i]; 1352 rs = &runstats[ts->groupid]; 1353 1354 if (is_backend) 1355 fio_server_send_ts(ts, rs); 1356 else if (output_format == FIO_OUTPUT_TERSE) 1357 show_thread_status_terse(ts, rs); 1358 else if (output_format == FIO_OUTPUT_JSON) { 1359 struct json_object *tmp = show_thread_status_json(ts, rs); 1360 json_array_add_value_object(array, tmp); 1361 } else 1362 show_thread_status_normal(ts, rs); 1363 } 1364 if (output_format == FIO_OUTPUT_JSON) { 1365 /* disk util stats, if any */ 1366 show_disk_util(1, root); 1367 1368 show_idle_prof_stats(FIO_OUTPUT_JSON, root); 1369 1370 json_print_object(root); 1371 log_info("\n"); 1372 json_free_object(root); 1373 } 1374 1375 for (i = 0; i < groupid + 1; i++) { 1376 rs = &runstats[i]; 1377 1378 rs->groupid = i; 1379 if (is_backend) 1380 fio_server_send_gs(rs); 1381 else if (output_format == FIO_OUTPUT_NORMAL) 1382 show_group_stats(rs); 1383 } 1384 1385 if (is_backend) 1386 fio_server_send_du(); 1387 else if (output_format == FIO_OUTPUT_NORMAL) { 1388 show_disk_util(0, NULL); 1389 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL); 1390 } 1391 1392 if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) { 1393 log_info("\nAdditional Terse Output:\n"); 1394 1395 for (i = 0; i < nr_ts; i++) { 1396 ts = &threadstats[i]; 1397 rs = &runstats[ts->groupid]; 1398 show_thread_status_terse(ts, rs); 1399 } 1400 } 1401 1402 log_info_flush(); 1403 free(runstats); 1404 free(threadstats); 1405 } 1406 1407 void show_run_stats(void) 1408 { 1409 fio_mutex_down(stat_mutex); 1410 __show_run_stats(); 1411 fio_mutex_up(stat_mutex); 1412 } 1413 1414 static void *__show_running_run_stats(void fio_unused *arg) 1415 { 1416 struct thread_data *td; 1417 unsigned long long *rt; 1418 struct timeval tv; 1419 int i; 1420 1421 rt = malloc(thread_number * sizeof(unsigned long long)); 1422 fio_gettime(&tv, NULL); 1423 1424 for_each_td(td, i) { 1425 rt[i] = mtime_since(&td->start, &tv); 1426 if (td_read(td) && td->io_bytes[DDIR_READ]) 1427 td->ts.runtime[DDIR_READ] += rt[i]; 1428 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1429 td->ts.runtime[DDIR_WRITE] += rt[i]; 1430 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1431 td->ts.runtime[DDIR_TRIM] += rt[i]; 1432 1433 td->update_rusage = 1; 1434 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; 1435 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; 1436 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; 1437 td->ts.total_run_time = mtime_since(&td->epoch, &tv); 1438 } 1439 1440 for_each_td(td, i) { 1441 if (td->rusage_sem) { 1442 td->update_rusage = 1; 1443 fio_mutex_down(td->rusage_sem); 1444 } 1445 td->update_rusage = 0; 1446 } 1447 1448 __show_run_stats(); 1449 1450 for_each_td(td, i) { 1451 if (td_read(td) && td->io_bytes[DDIR_READ]) 1452 td->ts.runtime[DDIR_READ] -= rt[i]; 1453 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1454 td->ts.runtime[DDIR_WRITE] -= rt[i]; 1455 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1456 td->ts.runtime[DDIR_TRIM] -= rt[i]; 1457 } 1458 1459 free(rt); 1460 fio_mutex_up(stat_mutex); 1461 return NULL; 1462 } 1463 1464 /* 1465 * Called from signal handler. It _should_ be safe to just run this inline 1466 * in the sig handler, but we should be disturbing the system less by just 1467 * creating a thread to do it. 1468 */ 1469 void show_running_run_stats(void) 1470 { 1471 pthread_t thread; 1472 1473 fio_mutex_down(stat_mutex); 1474 1475 if (!pthread_create(&thread, NULL, __show_running_run_stats, NULL)) { 1476 int err; 1477 1478 err = pthread_detach(thread); 1479 if (err) 1480 log_err("fio: DU thread detach failed: %s\n", strerror(err)); 1481 1482 return; 1483 } 1484 1485 fio_mutex_up(stat_mutex); 1486 } 1487 1488 static int status_interval_init; 1489 static struct timeval status_time; 1490 static int status_file_disabled; 1491 1492 #define FIO_STATUS_FILE "fio-dump-status" 1493 1494 static int check_status_file(void) 1495 { 1496 struct stat sb; 1497 const char *temp_dir; 1498 char fio_status_file_path[PATH_MAX]; 1499 1500 if (status_file_disabled) 1501 return 0; 1502 1503 temp_dir = getenv("TMPDIR"); 1504 if (temp_dir == NULL) { 1505 temp_dir = getenv("TEMP"); 1506 if (temp_dir && strlen(temp_dir) >= PATH_MAX) 1507 temp_dir = NULL; 1508 } 1509 if (temp_dir == NULL) 1510 temp_dir = "/tmp"; 1511 1512 snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE); 1513 1514 if (stat(fio_status_file_path, &sb)) 1515 return 0; 1516 1517 if (unlink(fio_status_file_path) < 0) { 1518 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path, 1519 strerror(errno)); 1520 log_err("fio: disabling status file updates\n"); 1521 status_file_disabled = 1; 1522 } 1523 1524 return 1; 1525 } 1526 1527 void check_for_running_stats(void) 1528 { 1529 if (status_interval) { 1530 if (!status_interval_init) { 1531 fio_gettime(&status_time, NULL); 1532 status_interval_init = 1; 1533 } else if (mtime_since_now(&status_time) >= status_interval) { 1534 show_running_run_stats(); 1535 fio_gettime(&status_time, NULL); 1536 return; 1537 } 1538 } 1539 if (check_status_file()) { 1540 show_running_run_stats(); 1541 return; 1542 } 1543 } 1544 1545 static inline void add_stat_sample(struct io_stat *is, unsigned long data) 1546 { 1547 double val = data; 1548 double delta; 1549 1550 if (data > is->max_val) 1551 is->max_val = data; 1552 if (data < is->min_val) 1553 is->min_val = data; 1554 1555 delta = val - is->mean.u.f; 1556 if (delta) { 1557 is->mean.u.f += delta / (is->samples + 1.0); 1558 is->S.u.f += delta * (val - is->mean.u.f); 1559 } 1560 1561 is->samples++; 1562 } 1563 1564 static void __add_log_sample(struct io_log *iolog, unsigned long val, 1565 enum fio_ddir ddir, unsigned int bs, 1566 unsigned long t) 1567 { 1568 const int nr_samples = iolog->nr_samples; 1569 1570 if (iolog->disabled) 1571 return; 1572 1573 if (!iolog->nr_samples) 1574 iolog->avg_last = t; 1575 1576 if (iolog->nr_samples == iolog->max_samples) { 1577 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 1578 void *new_log; 1579 1580 new_log = realloc(iolog->log, new_size); 1581 if (!new_log) { 1582 log_err("fio: failed extending iolog! Will stop logging.\n"); 1583 iolog->disabled = 1; 1584 return; 1585 } 1586 iolog->log = new_log; 1587 iolog->max_samples <<= 1; 1588 } 1589 1590 iolog->log[nr_samples].val = val; 1591 iolog->log[nr_samples].time = t; 1592 iolog->log[nr_samples].ddir = ddir; 1593 iolog->log[nr_samples].bs = bs; 1594 iolog->nr_samples++; 1595 } 1596 1597 static inline void reset_io_stat(struct io_stat *ios) 1598 { 1599 ios->max_val = ios->min_val = ios->samples = 0; 1600 ios->mean.u.f = ios->S.u.f = 0; 1601 } 1602 1603 void reset_io_stats(struct thread_data *td) 1604 { 1605 struct thread_stat *ts = &td->ts; 1606 int i, j; 1607 1608 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 1609 reset_io_stat(&ts->clat_stat[i]); 1610 reset_io_stat(&ts->slat_stat[i]); 1611 reset_io_stat(&ts->lat_stat[i]); 1612 reset_io_stat(&ts->bw_stat[i]); 1613 reset_io_stat(&ts->iops_stat[i]); 1614 1615 ts->io_bytes[i] = 0; 1616 ts->runtime[i] = 0; 1617 1618 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) 1619 ts->io_u_plat[i][j] = 0; 1620 } 1621 1622 for (i = 0; i < FIO_IO_U_MAP_NR; i++) { 1623 ts->io_u_map[i] = 0; 1624 ts->io_u_submit[i] = 0; 1625 ts->io_u_complete[i] = 0; 1626 ts->io_u_lat_u[i] = 0; 1627 ts->io_u_lat_m[i] = 0; 1628 ts->total_submit = 0; 1629 ts->total_complete = 0; 1630 } 1631 1632 for (i = 0; i < 3; i++) { 1633 ts->total_io_u[i] = 0; 1634 ts->short_io_u[i] = 0; 1635 } 1636 } 1637 1638 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed) 1639 { 1640 /* 1641 * Note an entry in the log. Use the mean from the logged samples, 1642 * making sure to properly round up. Only write a log entry if we 1643 * had actual samples done. 1644 */ 1645 if (iolog->avg_window[DDIR_READ].samples) { 1646 unsigned long mr; 1647 1648 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50; 1649 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed); 1650 } 1651 if (iolog->avg_window[DDIR_WRITE].samples) { 1652 unsigned long mw; 1653 1654 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50; 1655 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed); 1656 } 1657 if (iolog->avg_window[DDIR_TRIM].samples) { 1658 unsigned long mw; 1659 1660 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50; 1661 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed); 1662 } 1663 1664 reset_io_stat(&iolog->avg_window[DDIR_READ]); 1665 reset_io_stat(&iolog->avg_window[DDIR_WRITE]); 1666 reset_io_stat(&iolog->avg_window[DDIR_TRIM]); 1667 } 1668 1669 static void add_log_sample(struct thread_data *td, struct io_log *iolog, 1670 unsigned long val, enum fio_ddir ddir, 1671 unsigned int bs) 1672 { 1673 unsigned long elapsed, this_window; 1674 1675 if (!ddir_rw(ddir)) 1676 return; 1677 1678 elapsed = mtime_since_now(&td->epoch); 1679 1680 /* 1681 * If no time averaging, just add the log sample. 1682 */ 1683 if (!iolog->avg_msec) { 1684 __add_log_sample(iolog, val, ddir, bs, elapsed); 1685 return; 1686 } 1687 1688 /* 1689 * Add the sample. If the time period has passed, then 1690 * add that entry to the log and clear. 1691 */ 1692 add_stat_sample(&iolog->avg_window[ddir], val); 1693 1694 /* 1695 * If period hasn't passed, adding the above sample is all we 1696 * need to do. 1697 */ 1698 this_window = elapsed - iolog->avg_last; 1699 if (this_window < iolog->avg_msec) 1700 return; 1701 1702 _add_stat_to_log(iolog, elapsed); 1703 1704 iolog->avg_last = elapsed; 1705 } 1706 1707 void finalize_logs(struct thread_data *td) 1708 { 1709 unsigned long elapsed; 1710 1711 elapsed = mtime_since_now(&td->epoch); 1712 1713 if (td->clat_log) 1714 _add_stat_to_log(td->clat_log, elapsed); 1715 if (td->slat_log) 1716 _add_stat_to_log(td->slat_log, elapsed); 1717 if (td->lat_log) 1718 _add_stat_to_log(td->lat_log, elapsed); 1719 if (td->bw_log) 1720 _add_stat_to_log(td->bw_log, elapsed); 1721 if (td->iops_log) 1722 _add_stat_to_log(td->iops_log, elapsed); 1723 } 1724 1725 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 1726 { 1727 struct io_log *iolog; 1728 1729 if (!ddir_rw(ddir)) 1730 return; 1731 1732 iolog = agg_io_log[ddir]; 1733 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 1734 } 1735 1736 static void add_clat_percentile_sample(struct thread_stat *ts, 1737 unsigned long usec, enum fio_ddir ddir) 1738 { 1739 unsigned int idx = plat_val_to_idx(usec); 1740 assert(idx < FIO_IO_U_PLAT_NR); 1741 1742 ts->io_u_plat[ddir][idx]++; 1743 } 1744 1745 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 1746 unsigned long usec, unsigned int bs) 1747 { 1748 struct thread_stat *ts = &td->ts; 1749 1750 if (!ddir_rw(ddir)) 1751 return; 1752 1753 add_stat_sample(&ts->clat_stat[ddir], usec); 1754 1755 if (td->clat_log) 1756 add_log_sample(td, td->clat_log, usec, ddir, bs); 1757 1758 if (ts->clat_percentiles) 1759 add_clat_percentile_sample(ts, usec, ddir); 1760 } 1761 1762 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1763 unsigned long usec, unsigned int bs) 1764 { 1765 struct thread_stat *ts = &td->ts; 1766 1767 if (!ddir_rw(ddir)) 1768 return; 1769 1770 add_stat_sample(&ts->slat_stat[ddir], usec); 1771 1772 if (td->slat_log) 1773 add_log_sample(td, td->slat_log, usec, ddir, bs); 1774 } 1775 1776 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1777 unsigned long usec, unsigned int bs) 1778 { 1779 struct thread_stat *ts = &td->ts; 1780 1781 if (!ddir_rw(ddir)) 1782 return; 1783 1784 add_stat_sample(&ts->lat_stat[ddir], usec); 1785 1786 if (td->lat_log) 1787 add_log_sample(td, td->lat_log, usec, ddir, bs); 1788 } 1789 1790 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1791 struct timeval *t) 1792 { 1793 struct thread_stat *ts = &td->ts; 1794 unsigned long spent, rate; 1795 1796 if (!ddir_rw(ddir)) 1797 return; 1798 1799 spent = mtime_since(&td->bw_sample_time, t); 1800 if (spent < td->o.bw_avg_time) 1801 return; 1802 1803 /* 1804 * Compute both read and write rates for the interval. 1805 */ 1806 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1807 uint64_t delta; 1808 1809 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]; 1810 if (!delta) 1811 continue; /* No entries for interval */ 1812 1813 if (spent) 1814 rate = delta * 1000 / spent / 1024; 1815 else 1816 rate = 0; 1817 1818 add_stat_sample(&ts->bw_stat[ddir], rate); 1819 1820 if (td->bw_log) 1821 add_log_sample(td, td->bw_log, rate, ddir, bs); 1822 1823 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1824 } 1825 1826 fio_gettime(&td->bw_sample_time, NULL); 1827 } 1828 1829 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1830 struct timeval *t) 1831 { 1832 struct thread_stat *ts = &td->ts; 1833 unsigned long spent, iops; 1834 1835 if (!ddir_rw(ddir)) 1836 return; 1837 1838 spent = mtime_since(&td->iops_sample_time, t); 1839 if (spent < td->o.iops_avg_time) 1840 return; 1841 1842 /* 1843 * Compute both read and write rates for the interval. 1844 */ 1845 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1846 uint64_t delta; 1847 1848 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]; 1849 if (!delta) 1850 continue; /* No entries for interval */ 1851 1852 if (spent) 1853 iops = (delta * 1000) / spent; 1854 else 1855 iops = 0; 1856 1857 add_stat_sample(&ts->iops_stat[ddir], iops); 1858 1859 if (td->iops_log) 1860 add_log_sample(td, td->iops_log, iops, ddir, bs); 1861 1862 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1863 } 1864 1865 fio_gettime(&td->iops_sample_time, NULL); 1866 } 1867 1868 void stat_init(void) 1869 { 1870 stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED); 1871 } 1872 1873 void stat_exit(void) 1874 { 1875 /* 1876 * When we have the mutex, we know out-of-band access to it 1877 * have ended. 1878 */ 1879 fio_mutex_down(stat_mutex); 1880 fio_mutex_remove(stat_mutex); 1881 } 1882
