stat.c revision d686990a3c4bf23823d78c9b69196b24028491ee
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 17void update_rusage_stat(struct thread_data *td) 18{ 19 struct thread_stat *ts = &td->ts; 20 21 fio_getrusage(&td->ru_end); 22 ts->usr_time += mtime_since(&td->ru_start.ru_utime, 23 &td->ru_end.ru_utime); 24 ts->sys_time += mtime_since(&td->ru_start.ru_stime, 25 &td->ru_end.ru_stime); 26 ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw 27 - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw); 28 ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt; 29 ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt; 30 31 memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end)); 32} 33 34/* 35 * Given a latency, return the index of the corresponding bucket in 36 * the structure tracking percentiles. 37 * 38 * (1) find the group (and error bits) that the value (latency) 39 * belongs to by looking at its MSB. (2) find the bucket number in the 40 * group by looking at the index bits. 41 * 42 */ 43static unsigned int plat_val_to_idx(unsigned int val) 44{ 45 unsigned int msb, error_bits, base, offset, idx; 46 47 /* Find MSB starting from bit 0 */ 48 if (val == 0) 49 msb = 0; 50 else 51 msb = (sizeof(val)*8) - __builtin_clz(val) - 1; 52 53 /* 54 * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 55 * all bits of the sample as index 56 */ 57 if (msb <= FIO_IO_U_PLAT_BITS) 58 return val; 59 60 /* Compute the number of error bits to discard*/ 61 error_bits = msb - FIO_IO_U_PLAT_BITS; 62 63 /* Compute the number of buckets before the group */ 64 base = (error_bits + 1) << FIO_IO_U_PLAT_BITS; 65 66 /* 67 * Discard the error bits and apply the mask to find the 68 * index for the buckets in the group 69 */ 70 offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits); 71 72 /* Make sure the index does not exceed (array size - 1) */ 73 idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1)? 74 (base + offset) : (FIO_IO_U_PLAT_NR - 1); 75 76 return idx; 77} 78 79/* 80 * Convert the given index of the bucket array to the value 81 * represented by the bucket 82 */ 83static unsigned int plat_idx_to_val(unsigned int idx) 84{ 85 unsigned int error_bits, k, base; 86 87 assert(idx < FIO_IO_U_PLAT_NR); 88 89 /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use 90 * all bits of the sample as index */ 91 if (idx < (FIO_IO_U_PLAT_VAL << 1) ) 92 return idx; 93 94 /* Find the group and compute the minimum value of that group */ 95 error_bits = (idx >> FIO_IO_U_PLAT_BITS) -1; 96 base = 1 << (error_bits + FIO_IO_U_PLAT_BITS); 97 98 /* Find its bucket number of the group */ 99 k = idx % FIO_IO_U_PLAT_VAL; 100 101 /* Return the mean of the range of the bucket */ 102 return base + ((k + 0.5) * (1 << error_bits)); 103} 104 105static int double_cmp(const void *a, const void *b) 106{ 107 const fio_fp64_t fa = *(const fio_fp64_t *) a; 108 const fio_fp64_t fb = *(const fio_fp64_t *) b; 109 int cmp = 0; 110 111 if (fa.u.f > fb.u.f) 112 cmp = 1; 113 else if (fa.u.f < fb.u.f) 114 cmp = -1; 115 116 return cmp; 117} 118 119static unsigned int calc_clat_percentiles(unsigned int *io_u_plat, 120 unsigned long nr, fio_fp64_t *plist, 121 unsigned int **output, 122 unsigned int *maxv, 123 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 */ 184static 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 238out: 239 if (ovals) 240 free(ovals); 241} 242 243static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, 244 double *mean, double *dev) 245{ 246 double n = is->samples; 247 248 if (is->samples == 0) 249 return 0; 250 251 *min = is->min_val; 252 *max = is->max_val; 253 254 n = (double) is->samples; 255 *mean = is->mean.u.f; 256 257 if (n > 1.0) 258 *dev = sqrt(is->S.u.f / (n - 1.0)); 259 else 260 *dev = 0; 261 262 return 1; 263} 264 265void show_group_stats(struct group_run_stats *rs) 266{ 267 char *p1, *p2, *p3, *p4; 268 const char *ddir_str[] = { " READ", " WRITE" , " TRIM"}; 269 int i; 270 271 log_info("\nRun status group %d (all jobs):\n", rs->groupid); 272 273 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 274 const int i2p = is_power_of_2(rs->kb_base); 275 276 if (!rs->max_run[i]) 277 continue; 278 279 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8); 280 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base); 281 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base); 282 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base); 283 284 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s," 285 " mint=%llumsec, maxt=%llumsec\n", 286 rs->unified_rw_rep ? " MIXED" : ddir_str[i], 287 p1, p2, p3, p4, rs->min_run[i], rs->max_run[i]); 288 289 free(p1); 290 free(p2); 291 free(p3); 292 free(p4); 293 } 294} 295 296#define ts_total_io_u(ts) \ 297 ((ts)->total_io_u[DDIR_READ] + (ts)->total_io_u[DDIR_WRITE] +\ 298 (ts)->total_io_u[DDIR_TRIM]) 299 300static void stat_calc_dist(unsigned int *map, unsigned long total, 301 double *io_u_dist) 302{ 303 int i; 304 305 /* 306 * Do depth distribution calculations 307 */ 308 for (i = 0; i < FIO_IO_U_MAP_NR; i++) { 309 if (total) { 310 io_u_dist[i] = (double) map[i] / (double) total; 311 io_u_dist[i] *= 100.0; 312 if (io_u_dist[i] < 0.1 && map[i]) 313 io_u_dist[i] = 0.1; 314 } else 315 io_u_dist[i] = 0.0; 316 } 317} 318 319static void stat_calc_lat(struct thread_stat *ts, double *dst, 320 unsigned int *src, int nr) 321{ 322 unsigned long total = ts_total_io_u(ts); 323 int i; 324 325 /* 326 * Do latency distribution calculations 327 */ 328 for (i = 0; i < nr; i++) { 329 if (total) { 330 dst[i] = (double) src[i] / (double) total; 331 dst[i] *= 100.0; 332 if (dst[i] < 0.01 && src[i]) 333 dst[i] = 0.01; 334 } else 335 dst[i] = 0.0; 336 } 337} 338 339static void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat) 340{ 341 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR); 342} 343 344static void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat) 345{ 346 stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR); 347} 348 349static int usec_to_msec(unsigned long *min, unsigned long *max, double *mean, 350 double *dev) 351{ 352 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) { 353 *min /= 1000; 354 *max /= 1000; 355 *mean /= 1000.0; 356 *dev /= 1000.0; 357 return 0; 358 } 359 360 return 1; 361} 362 363static 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, ts->runtime[ddir]); 391 392 free(io_p); 393 free(bw_p); 394 free(iops_p); 395 396 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 397 const char *base = "(usec)"; 398 char *minp, *maxp; 399 400 if (!usec_to_msec(&min, &max, &mean, &dev)) 401 base = "(msec)"; 402 403 minp = num2str(min, 6, 1, 0, 0); 404 maxp = num2str(max, 6, 1, 0, 0); 405 406 log_info(" slat %s: min=%s, max=%s, avg=%5.02f," 407 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 408 409 free(minp); 410 free(maxp); 411 } 412 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 413 const char *base = "(usec)"; 414 char *minp, *maxp; 415 416 if (!usec_to_msec(&min, &max, &mean, &dev)) 417 base = "(msec)"; 418 419 minp = num2str(min, 6, 1, 0, 0); 420 maxp = num2str(max, 6, 1, 0, 0); 421 422 log_info(" clat %s: min=%s, max=%s, avg=%5.02f," 423 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 424 425 free(minp); 426 free(maxp); 427 } 428 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 429 const char *base = "(usec)"; 430 char *minp, *maxp; 431 432 if (!usec_to_msec(&min, &max, &mean, &dev)) 433 base = "(msec)"; 434 435 minp = num2str(min, 6, 1, 0, 0); 436 maxp = num2str(max, 6, 1, 0, 0); 437 438 log_info(" lat %s: min=%s, max=%s, avg=%5.02f," 439 " stdev=%5.02f\n", base, minp, maxp, mean, dev); 440 441 free(minp); 442 free(maxp); 443 } 444 if (ts->clat_percentiles) { 445 show_clat_percentiles(ts->io_u_plat[ddir], 446 ts->clat_stat[ddir].samples, 447 ts->percentile_list, 448 ts->percentile_precision); 449 } 450 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 451 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base; 452 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB"); 453 454 if (rs->unit_base == 1) { 455 min *= 8.0; 456 max *= 8.0; 457 mean *= 8.0; 458 dev *= 8.0; 459 } 460 461 if (rs->agg[ddir]) { 462 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 463 if (p_of_agg > 100.0) 464 p_of_agg = 100.0; 465 } 466 467 if (mean > fkb_base * fkb_base) { 468 min /= fkb_base; 469 max /= fkb_base; 470 mean /= fkb_base; 471 dev /= fkb_base; 472 bw_str = (rs->unit_base == 1 ? "Mbit" : "MB"); 473 } 474 475 log_info(" bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%," 476 " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max, 477 p_of_agg, mean, dev); 478 } 479} 480 481static int show_lat(double *io_u_lat, int nr, const char **ranges, 482 const char *msg) 483{ 484 int new_line = 1, i, line = 0, shown = 0; 485 486 for (i = 0; i < nr; i++) { 487 if (io_u_lat[i] <= 0.0) 488 continue; 489 shown = 1; 490 if (new_line) { 491 if (line) 492 log_info("\n"); 493 log_info(" lat (%s) : ", msg); 494 new_line = 0; 495 line = 0; 496 } 497 if (line) 498 log_info(", "); 499 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]); 500 line++; 501 if (line == 5) 502 new_line = 1; 503 } 504 505 if (shown) 506 log_info("\n"); 507 508 return shown; 509} 510 511static void show_lat_u(double *io_u_lat_u) 512{ 513 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 514 "250=", "500=", "750=", "1000=", }; 515 516 show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec"); 517} 518 519static void show_lat_m(double *io_u_lat_m) 520{ 521 const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=", 522 "250=", "500=", "750=", "1000=", "2000=", 523 ">=2000=", }; 524 525 show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec"); 526} 527 528static void show_latencies(double *io_u_lat_u, double *io_u_lat_m) 529{ 530 show_lat_u(io_u_lat_u); 531 show_lat_m(io_u_lat_m); 532} 533 534void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs) 535{ 536 double usr_cpu, sys_cpu; 537 unsigned long runtime; 538 double io_u_dist[FIO_IO_U_MAP_NR]; 539 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 540 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 541 time_t time_p; 542 char time_buf[64]; 543 544 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] + 545 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] + 546 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM])) 547 return; 548 549 time(&time_p); 550 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf)); 551 552 if (!ts->error) { 553 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s", 554 ts->name, ts->groupid, ts->members, 555 ts->error, (int) ts->pid, time_buf); 556 } else { 557 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s", 558 ts->name, ts->groupid, ts->members, 559 ts->error, ts->verror, (int) ts->pid, 560 time_buf); 561 } 562 563 if (strlen(ts->description)) 564 log_info(" Description : [%s]\n", ts->description); 565 566 if (ts->io_bytes[DDIR_READ]) 567 show_ddir_status(rs, ts, DDIR_READ); 568 if (ts->io_bytes[DDIR_WRITE]) 569 show_ddir_status(rs, ts, DDIR_WRITE); 570 if (ts->io_bytes[DDIR_TRIM]) 571 show_ddir_status(rs, ts, DDIR_TRIM); 572 573 stat_calc_lat_u(ts, io_u_lat_u); 574 stat_calc_lat_m(ts, io_u_lat_m); 575 show_latencies(io_u_lat_u, io_u_lat_m); 576 577 runtime = ts->total_run_time; 578 if (runtime) { 579 double runt = (double) runtime; 580 581 usr_cpu = (double) ts->usr_time * 100 / runt; 582 sys_cpu = (double) ts->sys_time * 100 / runt; 583 } else { 584 usr_cpu = 0; 585 sys_cpu = 0; 586 } 587 588 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu, majf=%lu," 589 " minf=%lu\n", usr_cpu, sys_cpu, ts->ctx, ts->majf, ts->minf); 590 591 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 592 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 593 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 594 io_u_dist[1], io_u_dist[2], 595 io_u_dist[3], io_u_dist[4], 596 io_u_dist[5], io_u_dist[6]); 597 598 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 599 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 600 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 601 io_u_dist[1], io_u_dist[2], 602 io_u_dist[3], io_u_dist[4], 603 io_u_dist[5], io_u_dist[6]); 604 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 605 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 606 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 607 io_u_dist[1], io_u_dist[2], 608 io_u_dist[3], io_u_dist[4], 609 io_u_dist[5], io_u_dist[6]); 610 log_info(" issued : total=r=%lu/w=%lu/d=%lu," 611 " short=r=%lu/w=%lu/d=%lu\n", 612 ts->total_io_u[0], ts->total_io_u[1], 613 ts->total_io_u[2], 614 ts->short_io_u[0], ts->short_io_u[1], 615 ts->short_io_u[2]); 616 if (ts->continue_on_error) { 617 log_info(" errors : total=%lu, first_error=%d/<%s>\n", 618 ts->total_err_count, 619 ts->first_error, 620 strerror(ts->first_error)); 621 } 622} 623 624static void show_ddir_status_terse(struct thread_stat *ts, 625 struct group_run_stats *rs, int ddir) 626{ 627 unsigned long min, max; 628 unsigned long long bw, iops; 629 unsigned int *ovals = NULL; 630 double mean, dev; 631 unsigned int len, minv, maxv; 632 int i; 633 634 assert(ddir_rw(ddir)); 635 636 iops = bw = 0; 637 if (ts->runtime[ddir]) { 638 uint64_t runt = ts->runtime[ddir]; 639 640 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 641 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 642 } 643 644 log_info(";%llu;%llu;%llu;%llu", ts->io_bytes[ddir] >> 10, bw, iops, 645 ts->runtime[ddir]); 646 647 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 648 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 649 else 650 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 651 652 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 653 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 654 else 655 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 656 657 if (ts->clat_percentiles) { 658 len = calc_clat_percentiles(ts->io_u_plat[ddir], 659 ts->clat_stat[ddir].samples, 660 ts->percentile_list, &ovals, &maxv, 661 &minv); 662 } else 663 len = 0; 664 665 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 666 if (i >= len) { 667 log_info(";0%%=0"); 668 continue; 669 } 670 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]); 671 } 672 673 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 674 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 675 else 676 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 677 678 if (ovals) 679 free(ovals); 680 681 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 682 double p_of_agg = 100.0; 683 684 if (rs->agg[ddir]) { 685 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 686 if (p_of_agg > 100.0) 687 p_of_agg = 100.0; 688 } 689 690 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 691 } else 692 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 693} 694 695static void add_ddir_status_json(struct thread_stat *ts, 696 struct group_run_stats *rs, int ddir, struct json_object *parent) 697{ 698 unsigned long min, max; 699 unsigned long long bw, iops; 700 unsigned int *ovals = NULL; 701 double mean, dev; 702 unsigned int len, minv, maxv; 703 int i; 704 const char *ddirname[] = {"read", "write", "trim"}; 705 struct json_object *dir_object, *tmp_object, *percentile_object; 706 char buf[120]; 707 double p_of_agg = 100.0; 708 709 assert(ddir_rw(ddir)); 710 711 if (ts->unified_rw_rep && ddir != DDIR_READ) 712 return; 713 714 dir_object = json_create_object(); 715 json_object_add_value_object(parent, 716 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object); 717 718 iops = bw = 0; 719 if (ts->runtime[ddir]) { 720 uint64_t runt = ts->runtime[ddir]; 721 722 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 723 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 724 } 725 726 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10); 727 json_object_add_value_int(dir_object, "bw", bw); 728 json_object_add_value_int(dir_object, "iops", iops); 729 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]); 730 731 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 732 min = max = 0; 733 mean = dev = 0.0; 734 } 735 tmp_object = json_create_object(); 736 json_object_add_value_object(dir_object, "slat", tmp_object); 737 json_object_add_value_int(tmp_object, "min", min); 738 json_object_add_value_int(tmp_object, "max", max); 739 json_object_add_value_float(tmp_object, "mean", mean); 740 json_object_add_value_float(tmp_object, "stddev", dev); 741 742 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 743 min = max = 0; 744 mean = dev = 0.0; 745 } 746 tmp_object = json_create_object(); 747 json_object_add_value_object(dir_object, "clat", tmp_object); 748 json_object_add_value_int(tmp_object, "min", min); 749 json_object_add_value_int(tmp_object, "max", max); 750 json_object_add_value_float(tmp_object, "mean", mean); 751 json_object_add_value_float(tmp_object, "stddev", dev); 752 753 if (ts->clat_percentiles) { 754 len = calc_clat_percentiles(ts->io_u_plat[ddir], 755 ts->clat_stat[ddir].samples, 756 ts->percentile_list, &ovals, &maxv, 757 &minv); 758 } else 759 len = 0; 760 761 percentile_object = json_create_object(); 762 json_object_add_value_object(tmp_object, "percentile", percentile_object); 763 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 764 if (i >= len) { 765 json_object_add_value_int(percentile_object, "0.00", 0); 766 continue; 767 } 768 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f); 769 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]); 770 } 771 772 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 773 min = max = 0; 774 mean = dev = 0.0; 775 } 776 tmp_object = json_create_object(); 777 json_object_add_value_object(dir_object, "lat", tmp_object); 778 json_object_add_value_int(tmp_object, "min", min); 779 json_object_add_value_int(tmp_object, "max", max); 780 json_object_add_value_float(tmp_object, "mean", mean); 781 json_object_add_value_float(tmp_object, "stddev", dev); 782 if (ovals) 783 free(ovals); 784 785 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 786 if (rs->agg[ddir]) { 787 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 788 if (p_of_agg > 100.0) 789 p_of_agg = 100.0; 790 } 791 } else { 792 min = max = 0; 793 p_of_agg = mean = dev = 0.0; 794 } 795 json_object_add_value_int(dir_object, "bw_min", min); 796 json_object_add_value_int(dir_object, "bw_max", max); 797 json_object_add_value_float(dir_object, "bw_agg", mean); 798 json_object_add_value_float(dir_object, "bw_mean", mean); 799 json_object_add_value_float(dir_object, "bw_dev", dev); 800} 801 802static void show_thread_status_terse_v2(struct thread_stat *ts, 803 struct group_run_stats *rs) 804{ 805 double io_u_dist[FIO_IO_U_MAP_NR]; 806 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 807 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 808 double usr_cpu, sys_cpu; 809 int i; 810 811 /* General Info */ 812 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error); 813 /* Log Read Status */ 814 show_ddir_status_terse(ts, rs, DDIR_READ); 815 /* Log Write Status */ 816 show_ddir_status_terse(ts, rs, DDIR_WRITE); 817 /* Log Trim Status */ 818 show_ddir_status_terse(ts, rs, DDIR_TRIM); 819 820 /* CPU Usage */ 821 if (ts->total_run_time) { 822 double runt = (double) ts->total_run_time; 823 824 usr_cpu = (double) ts->usr_time * 100 / runt; 825 sys_cpu = (double) ts->sys_time * 100 / runt; 826 } else { 827 usr_cpu = 0; 828 sys_cpu = 0; 829 } 830 831 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 832 ts->minf); 833 834 /* Calc % distribution of IO depths, usecond, msecond latency */ 835 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 836 stat_calc_lat_u(ts, io_u_lat_u); 837 stat_calc_lat_m(ts, io_u_lat_m); 838 839 /* Only show fixed 7 I/O depth levels*/ 840 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 841 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 842 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 843 844 /* Microsecond latency */ 845 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 846 log_info(";%3.2f%%", io_u_lat_u[i]); 847 /* Millisecond latency */ 848 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 849 log_info(";%3.2f%%", io_u_lat_m[i]); 850 /* Additional output if continue_on_error set - default off*/ 851 if (ts->continue_on_error) 852 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 853 log_info("\n"); 854 855 /* Additional output if description is set */ 856 if (ts->description) 857 log_info(";%s", ts->description); 858 859 log_info("\n"); 860} 861 862static void show_thread_status_terse_v3_v4(struct thread_stat *ts, 863 struct group_run_stats *rs, int ver) 864{ 865 double io_u_dist[FIO_IO_U_MAP_NR]; 866 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 867 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 868 double usr_cpu, sys_cpu; 869 int i; 870 871 /* General Info */ 872 log_info("%d;%s;%s;%d;%d", ver, fio_version_string, 873 ts->name, ts->groupid, ts->error); 874 /* Log Read Status */ 875 show_ddir_status_terse(ts, rs, DDIR_READ); 876 /* Log Write Status */ 877 show_ddir_status_terse(ts, rs, DDIR_WRITE); 878 /* Log Trim Status */ 879 if (ver == 4) 880 show_ddir_status_terse(ts, rs, DDIR_TRIM); 881 882 /* CPU Usage */ 883 if (ts->total_run_time) { 884 double runt = (double) ts->total_run_time; 885 886 usr_cpu = (double) ts->usr_time * 100 / runt; 887 sys_cpu = (double) ts->sys_time * 100 / runt; 888 } else { 889 usr_cpu = 0; 890 sys_cpu = 0; 891 } 892 893 log_info(";%f%%;%f%%;%lu;%lu;%lu", usr_cpu, sys_cpu, ts->ctx, ts->majf, 894 ts->minf); 895 896 /* Calc % distribution of IO depths, usecond, msecond latency */ 897 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 898 stat_calc_lat_u(ts, io_u_lat_u); 899 stat_calc_lat_m(ts, io_u_lat_m); 900 901 /* Only show fixed 7 I/O depth levels*/ 902 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 903 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 904 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 905 906 /* Microsecond latency */ 907 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 908 log_info(";%3.2f%%", io_u_lat_u[i]); 909 /* Millisecond latency */ 910 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 911 log_info(";%3.2f%%", io_u_lat_m[i]); 912 913 /* disk util stats, if any */ 914 show_disk_util(1, NULL); 915 916 /* Additional output if continue_on_error set - default off*/ 917 if (ts->continue_on_error) 918 log_info(";%lu;%d", ts->total_err_count, ts->first_error); 919 920 /* Additional output if description is set */ 921 if (strlen(ts->description)) 922 log_info(";%s", ts->description); 923 924 log_info("\n"); 925} 926 927static struct json_object *show_thread_status_json(struct thread_stat *ts, 928 struct group_run_stats *rs) 929{ 930 struct json_object *root, *tmp; 931 double io_u_dist[FIO_IO_U_MAP_NR]; 932 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 933 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 934 double usr_cpu, sys_cpu; 935 int i; 936 937 root = json_create_object(); 938 json_object_add_value_string(root, "jobname", ts->name); 939 json_object_add_value_int(root, "groupid", ts->groupid); 940 json_object_add_value_int(root, "error", ts->error); 941 942 add_ddir_status_json(ts, rs, DDIR_READ, root); 943 add_ddir_status_json(ts, rs, DDIR_WRITE, root); 944 add_ddir_status_json(ts, rs, DDIR_TRIM, root); 945 946 /* CPU Usage */ 947 if (ts->total_run_time) { 948 double runt = (double) ts->total_run_time; 949 950 usr_cpu = (double) ts->usr_time * 100 / runt; 951 sys_cpu = (double) ts->sys_time * 100 / runt; 952 } else { 953 usr_cpu = 0; 954 sys_cpu = 0; 955 } 956 json_object_add_value_float(root, "usr_cpu", usr_cpu); 957 json_object_add_value_float(root, "sys_cpu", sys_cpu); 958 json_object_add_value_int(root, "ctx", ts->ctx); 959 json_object_add_value_int(root, "majf", ts->majf); 960 json_object_add_value_int(root, "minf", ts->minf); 961 962 963 /* Calc % distribution of IO depths, usecond, msecond latency */ 964 stat_calc_dist(ts->io_u_map, ts_total_io_u(ts), io_u_dist); 965 stat_calc_lat_u(ts, io_u_lat_u); 966 stat_calc_lat_m(ts, io_u_lat_m); 967 968 tmp = json_create_object(); 969 json_object_add_value_object(root, "iodepth_level", tmp); 970 /* Only show fixed 7 I/O depth levels*/ 971 for (i = 0; i < 7; i++) { 972 char name[20]; 973 if (i < 6) 974 snprintf(name, 20, "%d", 1 << i); 975 else 976 snprintf(name, 20, ">=%d", 1 << i); 977 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]); 978 } 979 980 tmp = json_create_object(); 981 json_object_add_value_object(root, "latency_us", tmp); 982 /* Microsecond latency */ 983 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) { 984 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 985 "250", "500", "750", "1000", }; 986 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]); 987 } 988 /* Millisecond latency */ 989 tmp = json_create_object(); 990 json_object_add_value_object(root, "latency_ms", tmp); 991 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) { 992 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 993 "250", "500", "750", "1000", "2000", 994 ">=2000", }; 995 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]); 996 } 997 998 /* Additional output if continue_on_error set - default off*/ 999 if (ts->continue_on_error) { 1000 json_object_add_value_int(root, "total_err", ts->total_err_count); 1001 json_object_add_value_int(root, "total_err", ts->first_error); 1002 } 1003 1004 /* Additional output if description is set */ 1005 if (strlen(ts->description)) 1006 json_object_add_value_string(root, "desc", ts->description); 1007 1008 return root; 1009} 1010 1011static void show_thread_status_terse(struct thread_stat *ts, 1012 struct group_run_stats *rs) 1013{ 1014 if (terse_version == 2) 1015 show_thread_status_terse_v2(ts, rs); 1016 else if (terse_version == 3 || terse_version == 4) 1017 show_thread_status_terse_v3_v4(ts, rs, terse_version); 1018 else 1019 log_err("fio: bad terse version!? %d\n", terse_version); 1020} 1021 1022static 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 1057void 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 1077void 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 1148void 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 1157void 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 1172void 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); 1237 if (td->o.description) 1238 strncpy(ts->description, td->o.description, 1239 FIO_JOBNAME_SIZE); 1240 else 1241 memset(ts->description, 0, FIO_JOBNAME_SIZE); 1242 1243 ts->groupid = td->groupid; 1244 1245 /* 1246 * first pid in group, not very useful... 1247 */ 1248 ts->pid = td->pid; 1249 1250 ts->kb_base = td->o.kb_base; 1251 ts->unit_base = td->o.unit_base; 1252 ts->unified_rw_rep = td->o.unified_rw_rep; 1253 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 1254 log_info("fio: kb_base differs for jobs in group, using" 1255 " %u as the base\n", ts->kb_base); 1256 kb_base_warned = 1; 1257 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) { 1258 log_info("fio: unit_base differs for jobs in group, using" 1259 " %u as the base\n", ts->unit_base); 1260 unit_base_warned = 1; 1261 } 1262 1263 ts->continue_on_error = td->o.continue_on_error; 1264 ts->total_err_count += td->total_err_count; 1265 ts->first_error = td->first_error; 1266 if (!ts->error) { 1267 if (!td->error && td->o.continue_on_error && 1268 td->first_error) { 1269 ts->error = td->first_error; 1270 strcpy(ts->verror, td->verror); 1271 } else if (td->error) { 1272 ts->error = td->error; 1273 strcpy(ts->verror, td->verror); 1274 } 1275 } 1276 1277 sum_thread_stats(ts, &td->ts, idx); 1278 } 1279 1280 for (i = 0; i < nr_ts; i++) { 1281 unsigned long long bw; 1282 1283 ts = &threadstats[i]; 1284 rs = &runstats[ts->groupid]; 1285 rs->kb_base = ts->kb_base; 1286 rs->unit_base = ts->unit_base; 1287 rs->unified_rw_rep += ts->unified_rw_rep; 1288 1289 for (j = 0; j < DDIR_RWDIR_CNT; j++) { 1290 if (!ts->runtime[j]) 1291 continue; 1292 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 1293 rs->min_run[j] = ts->runtime[j]; 1294 if (ts->runtime[j] > rs->max_run[j]) 1295 rs->max_run[j] = ts->runtime[j]; 1296 1297 bw = 0; 1298 if (ts->runtime[j]) { 1299 unsigned long runt = ts->runtime[j]; 1300 unsigned long long kb; 1301 1302 kb = ts->io_bytes[j] / rs->kb_base; 1303 bw = kb * 1000 / runt; 1304 } 1305 if (bw < rs->min_bw[j]) 1306 rs->min_bw[j] = bw; 1307 if (bw > rs->max_bw[j]) 1308 rs->max_bw[j] = bw; 1309 1310 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 1311 } 1312 } 1313 1314 for (i = 0; i < groupid + 1; i++) { 1315 int ddir; 1316 1317 rs = &runstats[i]; 1318 1319 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) { 1320 if (rs->max_run[ddir]) 1321 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) / 1322 rs->max_run[ddir]; 1323 } 1324 } 1325 1326 /* 1327 * don't overwrite last signal output 1328 */ 1329 if (output_format == FIO_OUTPUT_NORMAL) 1330 log_info("\n"); 1331 else if (output_format == FIO_OUTPUT_JSON) { 1332 root = json_create_object(); 1333 json_object_add_value_string(root, "fio version", fio_version_string); 1334 array = json_create_array(); 1335 json_object_add_value_array(root, "jobs", array); 1336 } 1337 1338 for (i = 0; i < nr_ts; i++) { 1339 ts = &threadstats[i]; 1340 rs = &runstats[ts->groupid]; 1341 1342 if (is_backend) 1343 fio_server_send_ts(ts, rs); 1344 else if (output_format == FIO_OUTPUT_TERSE) 1345 show_thread_status_terse(ts, rs); 1346 else if (output_format == FIO_OUTPUT_JSON) { 1347 struct json_object *tmp = show_thread_status_json(ts, rs); 1348 json_array_add_value_object(array, tmp); 1349 } else 1350 show_thread_status(ts, rs); 1351 } 1352 if (output_format == FIO_OUTPUT_JSON) { 1353 /* disk util stats, if any */ 1354 show_disk_util(1, root); 1355 1356 show_idle_prof_stats(FIO_OUTPUT_JSON, root); 1357 1358 json_print_object(root); 1359 log_info("\n"); 1360 json_free_object(root); 1361 } 1362 1363 for (i = 0; i < groupid + 1; i++) { 1364 rs = &runstats[i]; 1365 1366 rs->groupid = i; 1367 if (is_backend) 1368 fio_server_send_gs(rs); 1369 else if (output_format == FIO_OUTPUT_NORMAL) 1370 show_group_stats(rs); 1371 } 1372 1373 if (is_backend) 1374 fio_server_send_du(); 1375 else if (output_format == FIO_OUTPUT_NORMAL) { 1376 show_disk_util(0, NULL); 1377 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL); 1378 } 1379 1380 free(runstats); 1381 free(threadstats); 1382} 1383 1384static void *__show_running_run_stats(void *arg) 1385{ 1386 struct thread_data *td; 1387 unsigned long long *rt; 1388 struct timeval tv; 1389 int i; 1390 1391 rt = malloc(thread_number * sizeof(unsigned long long)); 1392 fio_gettime(&tv, NULL); 1393 1394 for_each_td(td, i) { 1395 rt[i] = mtime_since(&td->start, &tv); 1396 if (td_read(td) && td->io_bytes[DDIR_READ]) 1397 td->ts.runtime[DDIR_READ] += rt[i]; 1398 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1399 td->ts.runtime[DDIR_WRITE] += rt[i]; 1400 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1401 td->ts.runtime[DDIR_TRIM] += rt[i]; 1402 1403 td->update_rusage = 1; 1404 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; 1405 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; 1406 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; 1407 td->ts.total_run_time = mtime_since(&td->epoch, &tv); 1408 } 1409 1410 for_each_td(td, i) { 1411 if (td->rusage_sem) { 1412 td->update_rusage = 1; 1413 fio_mutex_down(td->rusage_sem); 1414 } 1415 td->update_rusage = 0; 1416 } 1417 1418 show_run_stats(); 1419 1420 for_each_td(td, i) { 1421 if (td_read(td) && td->io_bytes[DDIR_READ]) 1422 td->ts.runtime[DDIR_READ] -= rt[i]; 1423 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1424 td->ts.runtime[DDIR_WRITE] -= rt[i]; 1425 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1426 td->ts.runtime[DDIR_TRIM] -= rt[i]; 1427 } 1428 1429 free(rt); 1430 return NULL; 1431} 1432 1433/* 1434 * Called from signal handler. It _should_ be safe to just run this inline 1435 * in the sig handler, but we should be disturbing the system less by just 1436 * creating a thread to do it. 1437 */ 1438void show_running_run_stats(void) 1439{ 1440 pthread_t thread; 1441 1442 pthread_create(&thread, NULL, __show_running_run_stats, NULL); 1443 pthread_detach(thread); 1444} 1445 1446static inline void add_stat_sample(struct io_stat *is, unsigned long data) 1447{ 1448 double val = data; 1449 double delta; 1450 1451 if (data > is->max_val) 1452 is->max_val = data; 1453 if (data < is->min_val) 1454 is->min_val = data; 1455 1456 delta = val - is->mean.u.f; 1457 if (delta) { 1458 is->mean.u.f += delta / (is->samples + 1.0); 1459 is->S.u.f += delta * (val - is->mean.u.f); 1460 } 1461 1462 is->samples++; 1463} 1464 1465static void __add_log_sample(struct io_log *iolog, unsigned long val, 1466 enum fio_ddir ddir, unsigned int bs, 1467 unsigned long t) 1468{ 1469 const int nr_samples = iolog->nr_samples; 1470 1471 if (!iolog->nr_samples) 1472 iolog->avg_last = t; 1473 1474 if (iolog->nr_samples == iolog->max_samples) { 1475 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 1476 1477 iolog->log = realloc(iolog->log, new_size); 1478 iolog->max_samples <<= 1; 1479 } 1480 1481 iolog->log[nr_samples].val = val; 1482 iolog->log[nr_samples].time = t; 1483 iolog->log[nr_samples].ddir = ddir; 1484 iolog->log[nr_samples].bs = bs; 1485 iolog->nr_samples++; 1486} 1487 1488static inline void reset_io_stat(struct io_stat *ios) 1489{ 1490 ios->max_val = ios->min_val = ios->samples = 0; 1491 ios->mean.u.f = ios->S.u.f = 0; 1492} 1493 1494static void add_log_sample(struct thread_data *td, struct io_log *iolog, 1495 unsigned long val, enum fio_ddir ddir, 1496 unsigned int bs) 1497{ 1498 unsigned long elapsed, this_window; 1499 1500 if (!ddir_rw(ddir)) 1501 return; 1502 1503 elapsed = mtime_since_now(&td->epoch); 1504 1505 /* 1506 * If no time averaging, just add the log sample. 1507 */ 1508 if (!iolog->avg_msec) { 1509 __add_log_sample(iolog, val, ddir, bs, elapsed); 1510 return; 1511 } 1512 1513 /* 1514 * Add the sample. If the time period has passed, then 1515 * add that entry to the log and clear. 1516 */ 1517 add_stat_sample(&iolog->avg_window[ddir], val); 1518 1519 /* 1520 * If period hasn't passed, adding the above sample is all we 1521 * need to do. 1522 */ 1523 this_window = elapsed - iolog->avg_last; 1524 if (this_window < iolog->avg_msec) 1525 return; 1526 1527 /* 1528 * Note an entry in the log. Use the mean from the logged samples, 1529 * making sure to properly round up. Only write a log entry if we 1530 * had actual samples done. 1531 */ 1532 if (iolog->avg_window[DDIR_READ].samples) { 1533 unsigned long mr; 1534 1535 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50; 1536 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed); 1537 } 1538 if (iolog->avg_window[DDIR_WRITE].samples) { 1539 unsigned long mw; 1540 1541 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50; 1542 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed); 1543 } 1544 if (iolog->avg_window[DDIR_TRIM].samples) { 1545 unsigned long mw; 1546 1547 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50; 1548 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed); 1549 } 1550 1551 1552 reset_io_stat(&iolog->avg_window[DDIR_READ]); 1553 reset_io_stat(&iolog->avg_window[DDIR_WRITE]); 1554 reset_io_stat(&iolog->avg_window[DDIR_TRIM]); 1555 iolog->avg_last = elapsed; 1556} 1557 1558void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 1559{ 1560 struct io_log *iolog; 1561 1562 if (!ddir_rw(ddir)) 1563 return; 1564 1565 iolog = agg_io_log[ddir]; 1566 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 1567} 1568 1569static void add_clat_percentile_sample(struct thread_stat *ts, 1570 unsigned long usec, enum fio_ddir ddir) 1571{ 1572 unsigned int idx = plat_val_to_idx(usec); 1573 assert(idx < FIO_IO_U_PLAT_NR); 1574 1575 ts->io_u_plat[ddir][idx]++; 1576} 1577 1578void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 1579 unsigned long usec, unsigned int bs) 1580{ 1581 struct thread_stat *ts = &td->ts; 1582 1583 if (!ddir_rw(ddir)) 1584 return; 1585 1586 add_stat_sample(&ts->clat_stat[ddir], usec); 1587 1588 if (td->clat_log) 1589 add_log_sample(td, td->clat_log, usec, ddir, bs); 1590 1591 if (ts->clat_percentiles) 1592 add_clat_percentile_sample(ts, usec, ddir); 1593} 1594 1595void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1596 unsigned long usec, unsigned int bs) 1597{ 1598 struct thread_stat *ts = &td->ts; 1599 1600 if (!ddir_rw(ddir)) 1601 return; 1602 1603 add_stat_sample(&ts->slat_stat[ddir], usec); 1604 1605 if (td->slat_log) 1606 add_log_sample(td, td->slat_log, usec, ddir, bs); 1607} 1608 1609void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1610 unsigned long usec, unsigned int bs) 1611{ 1612 struct thread_stat *ts = &td->ts; 1613 1614 if (!ddir_rw(ddir)) 1615 return; 1616 1617 add_stat_sample(&ts->lat_stat[ddir], usec); 1618 1619 if (td->lat_log) 1620 add_log_sample(td, td->lat_log, usec, ddir, bs); 1621} 1622 1623void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1624 struct timeval *t) 1625{ 1626 struct thread_stat *ts = &td->ts; 1627 unsigned long spent, rate; 1628 1629 if (!ddir_rw(ddir)) 1630 return; 1631 1632 spent = mtime_since(&td->bw_sample_time, t); 1633 if (spent < td->o.bw_avg_time) 1634 return; 1635 1636 /* 1637 * Compute both read and write rates for the interval. 1638 */ 1639 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1640 uint64_t delta; 1641 1642 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]; 1643 if (!delta) 1644 continue; /* No entries for interval */ 1645 1646 rate = delta * 1000 / spent / 1024; 1647 add_stat_sample(&ts->bw_stat[ddir], rate); 1648 1649 if (td->bw_log) 1650 add_log_sample(td, td->bw_log, rate, ddir, bs); 1651 1652 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1653 } 1654 1655 fio_gettime(&td->bw_sample_time, NULL); 1656} 1657 1658void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, 1659 struct timeval *t) 1660{ 1661 struct thread_stat *ts = &td->ts; 1662 unsigned long spent, iops; 1663 1664 if (!ddir_rw(ddir)) 1665 return; 1666 1667 spent = mtime_since(&td->iops_sample_time, t); 1668 if (spent < td->o.iops_avg_time) 1669 return; 1670 1671 /* 1672 * Compute both read and write rates for the interval. 1673 */ 1674 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1675 uint64_t delta; 1676 1677 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]; 1678 if (!delta) 1679 continue; /* No entries for interval */ 1680 1681 iops = (delta * 1000) / spent; 1682 add_stat_sample(&ts->iops_stat[ddir], iops); 1683 1684 if (td->iops_log) 1685 add_log_sample(td, td->iops_log, iops, ddir, 0); 1686 1687 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1688 } 1689 1690 fio_gettime(&td->iops_sample_time, NULL); 1691} 1692