stat.c revision fdd5f15f8231e8c91c4deff22e630a34addd0fef
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 119unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 120 fio_fp64_t *plist, unsigned int **output, 121 unsigned int *maxv, unsigned int *minv) 122{ 123 unsigned long sum = 0; 124 unsigned int len, i, j = 0; 125 unsigned int oval_len = 0; 126 unsigned int *ovals = NULL; 127 int is_last; 128 129 *minv = -1U; 130 *maxv = 0; 131 132 len = 0; 133 while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0) 134 len++; 135 136 if (!len) 137 return 0; 138 139 /* 140 * Sort the percentile list. Note that it may already be sorted if 141 * we are using the default values, but since it's a short list this 142 * isn't a worry. Also note that this does not work for NaN values. 143 */ 144 if (len > 1) 145 qsort((void *)plist, len, sizeof(plist[0]), double_cmp); 146 147 /* 148 * Calculate bucket values, note down max and min values 149 */ 150 is_last = 0; 151 for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) { 152 sum += io_u_plat[i]; 153 while (sum >= (plist[j].u.f / 100.0 * nr)) { 154 assert(plist[j].u.f <= 100.0); 155 156 if (j == oval_len) { 157 oval_len += 100; 158 ovals = realloc(ovals, oval_len * sizeof(unsigned int)); 159 } 160 161 ovals[j] = plat_idx_to_val(i); 162 if (ovals[j] < *minv) 163 *minv = ovals[j]; 164 if (ovals[j] > *maxv) 165 *maxv = ovals[j]; 166 167 is_last = (j == len - 1); 168 if (is_last) 169 break; 170 171 j++; 172 } 173 } 174 175 *output = ovals; 176 return len; 177} 178 179/* 180 * Find and display the p-th percentile of clat 181 */ 182static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, 183 fio_fp64_t *plist, unsigned int precision) 184{ 185 unsigned int len, j = 0, minv, maxv; 186 unsigned int *ovals; 187 int is_last, per_line, scale_down; 188 char fmt[32]; 189 190 len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv); 191 if (!len) 192 goto out; 193 194 /* 195 * We default to usecs, but if the value range is such that we 196 * should scale down to msecs, do that. 197 */ 198 if (minv > 2000 && maxv > 99999) { 199 scale_down = 1; 200 log_info(" clat percentiles (msec):\n |"); 201 } else { 202 scale_down = 0; 203 log_info(" clat percentiles (usec):\n |"); 204 } 205 206 snprintf(fmt, sizeof(fmt), "%%1.%uf", precision); 207 per_line = (80 - 7) / (precision + 14); 208 209 for (j = 0; j < len; j++) { 210 char fbuf[16], *ptr = fbuf; 211 212 /* for formatting */ 213 if (j != 0 && (j % per_line) == 0) 214 log_info(" |"); 215 216 /* end of the list */ 217 is_last = (j == len - 1); 218 219 if (plist[j].u.f < 10.0) 220 ptr += sprintf(fbuf, " "); 221 222 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f); 223 224 if (scale_down) 225 ovals[j] = (ovals[j] + 999) / 1000; 226 227 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ','); 228 229 if (is_last) 230 break; 231 232 if ((j % per_line) == per_line - 1) /* for formatting */ 233 log_info("\n"); 234 } 235 236out: 237 if (ovals) 238 free(ovals); 239} 240 241int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, 242 double *mean, double *dev) 243{ 244 double n = is->samples; 245 246 if (is->samples == 0) 247 return 0; 248 249 *min = is->min_val; 250 *max = is->max_val; 251 252 n = (double) is->samples; 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 263void 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 296void 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 314static 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 334void 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 339void 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 344static 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 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, 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 441static 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 471static 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 479static 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 488static 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 500void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs) 501{ 502 double usr_cpu, sys_cpu; 503 unsigned long runtime; 504 double io_u_dist[FIO_IO_U_MAP_NR]; 505 time_t time_p; 506 char time_buf[64]; 507 508 if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] + 509 ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] + 510 ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM])) 511 return; 512 513 time(&time_p); 514 os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf)); 515 516 if (!ts->error) { 517 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s", 518 ts->name, ts->groupid, ts->members, 519 ts->error, (int) ts->pid, time_buf); 520 } else { 521 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s", 522 ts->name, ts->groupid, ts->members, 523 ts->error, ts->verror, (int) ts->pid, 524 time_buf); 525 } 526 527 if (strlen(ts->description)) 528 log_info(" Description : [%s]\n", ts->description); 529 530 if (ts->io_bytes[DDIR_READ]) 531 show_ddir_status(rs, ts, DDIR_READ); 532 if (ts->io_bytes[DDIR_WRITE]) 533 show_ddir_status(rs, ts, DDIR_WRITE); 534 if (ts->io_bytes[DDIR_TRIM]) 535 show_ddir_status(rs, ts, DDIR_TRIM); 536 537 show_latencies(ts); 538 539 runtime = ts->total_run_time; 540 if (runtime) { 541 double runt = (double) runtime; 542 543 usr_cpu = (double) ts->usr_time * 100 / runt; 544 sys_cpu = (double) ts->sys_time * 100 / runt; 545 } else { 546 usr_cpu = 0; 547 sys_cpu = 0; 548 } 549 550 log_info(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu," 551 " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu, 552 (unsigned long long) ts->ctx, 553 (unsigned long long) ts->majf, 554 (unsigned long long) ts->minf); 555 556 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 557 log_info(" IO depths : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%," 558 " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 559 io_u_dist[1], io_u_dist[2], 560 io_u_dist[3], io_u_dist[4], 561 io_u_dist[5], io_u_dist[6]); 562 563 stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist); 564 log_info(" submit : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 565 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 566 io_u_dist[1], io_u_dist[2], 567 io_u_dist[3], io_u_dist[4], 568 io_u_dist[5], io_u_dist[6]); 569 stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist); 570 log_info(" complete : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%," 571 " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0], 572 io_u_dist[1], io_u_dist[2], 573 io_u_dist[3], io_u_dist[4], 574 io_u_dist[5], io_u_dist[6]); 575 log_info(" issued : total=r=%llu/w=%llu/d=%llu," 576 " short=r=%llu/w=%llu/d=%llu\n", 577 (unsigned long long) ts->total_io_u[0], 578 (unsigned long long) ts->total_io_u[1], 579 (unsigned long long) ts->total_io_u[2], 580 (unsigned long long) ts->short_io_u[0], 581 (unsigned long long) ts->short_io_u[1], 582 (unsigned long long) ts->short_io_u[2]); 583 if (ts->continue_on_error) { 584 log_info(" errors : total=%llu, first_error=%d/<%s>\n", 585 (unsigned long long)ts->total_err_count, 586 ts->first_error, 587 strerror(ts->first_error)); 588 } 589} 590 591static void show_ddir_status_terse(struct thread_stat *ts, 592 struct group_run_stats *rs, int ddir) 593{ 594 unsigned long min, max; 595 unsigned long long bw, iops; 596 unsigned int *ovals = NULL; 597 double mean, dev; 598 unsigned int len, minv, maxv; 599 int i; 600 601 assert(ddir_rw(ddir)); 602 603 iops = bw = 0; 604 if (ts->runtime[ddir]) { 605 uint64_t runt = ts->runtime[ddir]; 606 607 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 608 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 609 } 610 611 log_info(";%llu;%llu;%llu;%llu", 612 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops, 613 (unsigned long long) ts->runtime[ddir]); 614 615 if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) 616 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 617 else 618 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 619 620 if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) 621 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 622 else 623 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 624 625 if (ts->clat_percentiles) { 626 len = calc_clat_percentiles(ts->io_u_plat[ddir], 627 ts->clat_stat[ddir].samples, 628 ts->percentile_list, &ovals, &maxv, 629 &minv); 630 } else 631 len = 0; 632 633 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 634 if (i >= len) { 635 log_info(";0%%=0"); 636 continue; 637 } 638 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]); 639 } 640 641 if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) 642 log_info(";%lu;%lu;%f;%f", min, max, mean, dev); 643 else 644 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0); 645 646 if (ovals) 647 free(ovals); 648 649 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 650 double p_of_agg = 100.0; 651 652 if (rs->agg[ddir]) { 653 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 654 if (p_of_agg > 100.0) 655 p_of_agg = 100.0; 656 } 657 658 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev); 659 } else 660 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0); 661} 662 663static void add_ddir_status_json(struct thread_stat *ts, 664 struct group_run_stats *rs, int ddir, struct json_object *parent) 665{ 666 unsigned long min, max; 667 unsigned long long bw, iops; 668 unsigned int *ovals = NULL; 669 double mean, dev; 670 unsigned int len, minv, maxv; 671 int i; 672 const char *ddirname[] = {"read", "write", "trim"}; 673 struct json_object *dir_object, *tmp_object, *percentile_object; 674 char buf[120]; 675 double p_of_agg = 100.0; 676 677 assert(ddir_rw(ddir)); 678 679 if (ts->unified_rw_rep && ddir != DDIR_READ) 680 return; 681 682 dir_object = json_create_object(); 683 json_object_add_value_object(parent, 684 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object); 685 686 iops = bw = 0; 687 if (ts->runtime[ddir]) { 688 uint64_t runt = ts->runtime[ddir]; 689 690 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; 691 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt; 692 } 693 694 json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10); 695 json_object_add_value_int(dir_object, "bw", bw); 696 json_object_add_value_int(dir_object, "iops", iops); 697 json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]); 698 699 if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) { 700 min = max = 0; 701 mean = dev = 0.0; 702 } 703 tmp_object = json_create_object(); 704 json_object_add_value_object(dir_object, "slat", tmp_object); 705 json_object_add_value_int(tmp_object, "min", min); 706 json_object_add_value_int(tmp_object, "max", max); 707 json_object_add_value_float(tmp_object, "mean", mean); 708 json_object_add_value_float(tmp_object, "stddev", dev); 709 710 if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) { 711 min = max = 0; 712 mean = dev = 0.0; 713 } 714 tmp_object = json_create_object(); 715 json_object_add_value_object(dir_object, "clat", tmp_object); 716 json_object_add_value_int(tmp_object, "min", min); 717 json_object_add_value_int(tmp_object, "max", max); 718 json_object_add_value_float(tmp_object, "mean", mean); 719 json_object_add_value_float(tmp_object, "stddev", dev); 720 721 if (ts->clat_percentiles) { 722 len = calc_clat_percentiles(ts->io_u_plat[ddir], 723 ts->clat_stat[ddir].samples, 724 ts->percentile_list, &ovals, &maxv, 725 &minv); 726 } else 727 len = 0; 728 729 percentile_object = json_create_object(); 730 json_object_add_value_object(tmp_object, "percentile", percentile_object); 731 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) { 732 if (i >= len) { 733 json_object_add_value_int(percentile_object, "0.00", 0); 734 continue; 735 } 736 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f); 737 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]); 738 } 739 740 if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) { 741 min = max = 0; 742 mean = dev = 0.0; 743 } 744 tmp_object = json_create_object(); 745 json_object_add_value_object(dir_object, "lat", tmp_object); 746 json_object_add_value_int(tmp_object, "min", min); 747 json_object_add_value_int(tmp_object, "max", max); 748 json_object_add_value_float(tmp_object, "mean", mean); 749 json_object_add_value_float(tmp_object, "stddev", dev); 750 if (ovals) 751 free(ovals); 752 753 if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) { 754 if (rs->agg[ddir]) { 755 p_of_agg = mean * 100 / (double) rs->agg[ddir]; 756 if (p_of_agg > 100.0) 757 p_of_agg = 100.0; 758 } 759 } else { 760 min = max = 0; 761 p_of_agg = mean = dev = 0.0; 762 } 763 json_object_add_value_int(dir_object, "bw_min", min); 764 json_object_add_value_int(dir_object, "bw_max", max); 765 json_object_add_value_float(dir_object, "bw_agg", mean); 766 json_object_add_value_float(dir_object, "bw_mean", mean); 767 json_object_add_value_float(dir_object, "bw_dev", dev); 768} 769 770static void show_thread_status_terse_v2(struct thread_stat *ts, 771 struct group_run_stats *rs) 772{ 773 double io_u_dist[FIO_IO_U_MAP_NR]; 774 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 775 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 776 double usr_cpu, sys_cpu; 777 int i; 778 779 /* General Info */ 780 log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error); 781 /* Log Read Status */ 782 show_ddir_status_terse(ts, rs, DDIR_READ); 783 /* Log Write Status */ 784 show_ddir_status_terse(ts, rs, DDIR_WRITE); 785 /* Log Trim Status */ 786 show_ddir_status_terse(ts, rs, DDIR_TRIM); 787 788 /* CPU Usage */ 789 if (ts->total_run_time) { 790 double runt = (double) ts->total_run_time; 791 792 usr_cpu = (double) ts->usr_time * 100 / runt; 793 sys_cpu = (double) ts->sys_time * 100 / runt; 794 } else { 795 usr_cpu = 0; 796 sys_cpu = 0; 797 } 798 799 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, 800 (unsigned long long) ts->ctx, 801 (unsigned long long) ts->majf, 802 (unsigned long long) ts->minf); 803 804 /* Calc % distribution of IO depths, usecond, msecond latency */ 805 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 806 stat_calc_lat_u(ts, io_u_lat_u); 807 stat_calc_lat_m(ts, io_u_lat_m); 808 809 /* Only show fixed 7 I/O depth levels*/ 810 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 811 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 812 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 813 814 /* Microsecond latency */ 815 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 816 log_info(";%3.2f%%", io_u_lat_u[i]); 817 /* Millisecond latency */ 818 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 819 log_info(";%3.2f%%", io_u_lat_m[i]); 820 /* Additional output if continue_on_error set - default off*/ 821 if (ts->continue_on_error) 822 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); 823 log_info("\n"); 824 825 /* Additional output if description is set */ 826 if (ts->description) 827 log_info(";%s", ts->description); 828 829 log_info("\n"); 830} 831 832static void show_thread_status_terse_v3_v4(struct thread_stat *ts, 833 struct group_run_stats *rs, int ver) 834{ 835 double io_u_dist[FIO_IO_U_MAP_NR]; 836 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 837 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 838 double usr_cpu, sys_cpu; 839 int i; 840 841 /* General Info */ 842 log_info("%d;%s;%s;%d;%d", ver, fio_version_string, 843 ts->name, ts->groupid, ts->error); 844 /* Log Read Status */ 845 show_ddir_status_terse(ts, rs, DDIR_READ); 846 /* Log Write Status */ 847 show_ddir_status_terse(ts, rs, DDIR_WRITE); 848 /* Log Trim Status */ 849 if (ver == 4) 850 show_ddir_status_terse(ts, rs, DDIR_TRIM); 851 852 /* CPU Usage */ 853 if (ts->total_run_time) { 854 double runt = (double) ts->total_run_time; 855 856 usr_cpu = (double) ts->usr_time * 100 / runt; 857 sys_cpu = (double) ts->sys_time * 100 / runt; 858 } else { 859 usr_cpu = 0; 860 sys_cpu = 0; 861 } 862 863 log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu, 864 (unsigned long long) ts->ctx, 865 (unsigned long long) ts->majf, 866 (unsigned long long) ts->minf); 867 868 /* Calc % distribution of IO depths, usecond, msecond latency */ 869 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 870 stat_calc_lat_u(ts, io_u_lat_u); 871 stat_calc_lat_m(ts, io_u_lat_m); 872 873 /* Only show fixed 7 I/O depth levels*/ 874 log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%", 875 io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3], 876 io_u_dist[4], io_u_dist[5], io_u_dist[6]); 877 878 /* Microsecond latency */ 879 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) 880 log_info(";%3.2f%%", io_u_lat_u[i]); 881 /* Millisecond latency */ 882 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) 883 log_info(";%3.2f%%", io_u_lat_m[i]); 884 885 /* disk util stats, if any */ 886 show_disk_util(1, NULL); 887 888 /* Additional output if continue_on_error set - default off*/ 889 if (ts->continue_on_error) 890 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error); 891 892 /* Additional output if description is set */ 893 if (strlen(ts->description)) 894 log_info(";%s", ts->description); 895 896 log_info("\n"); 897} 898 899static struct json_object *show_thread_status_json(struct thread_stat *ts, 900 struct group_run_stats *rs) 901{ 902 struct json_object *root, *tmp; 903 double io_u_dist[FIO_IO_U_MAP_NR]; 904 double io_u_lat_u[FIO_IO_U_LAT_U_NR]; 905 double io_u_lat_m[FIO_IO_U_LAT_M_NR]; 906 double usr_cpu, sys_cpu; 907 int i; 908 909 root = json_create_object(); 910 json_object_add_value_string(root, "jobname", ts->name); 911 json_object_add_value_int(root, "groupid", ts->groupid); 912 json_object_add_value_int(root, "error", ts->error); 913 914 add_ddir_status_json(ts, rs, DDIR_READ, root); 915 add_ddir_status_json(ts, rs, DDIR_WRITE, root); 916 add_ddir_status_json(ts, rs, DDIR_TRIM, root); 917 918 /* CPU Usage */ 919 if (ts->total_run_time) { 920 double runt = (double) ts->total_run_time; 921 922 usr_cpu = (double) ts->usr_time * 100 / runt; 923 sys_cpu = (double) ts->sys_time * 100 / runt; 924 } else { 925 usr_cpu = 0; 926 sys_cpu = 0; 927 } 928 json_object_add_value_float(root, "usr_cpu", usr_cpu); 929 json_object_add_value_float(root, "sys_cpu", sys_cpu); 930 json_object_add_value_int(root, "ctx", ts->ctx); 931 json_object_add_value_int(root, "majf", ts->majf); 932 json_object_add_value_int(root, "minf", ts->minf); 933 934 935 /* Calc % distribution of IO depths, usecond, msecond latency */ 936 stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist); 937 stat_calc_lat_u(ts, io_u_lat_u); 938 stat_calc_lat_m(ts, io_u_lat_m); 939 940 tmp = json_create_object(); 941 json_object_add_value_object(root, "iodepth_level", tmp); 942 /* Only show fixed 7 I/O depth levels*/ 943 for (i = 0; i < 7; i++) { 944 char name[20]; 945 if (i < 6) 946 snprintf(name, 20, "%d", 1 << i); 947 else 948 snprintf(name, 20, ">=%d", 1 << i); 949 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]); 950 } 951 952 tmp = json_create_object(); 953 json_object_add_value_object(root, "latency_us", tmp); 954 /* Microsecond latency */ 955 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) { 956 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 957 "250", "500", "750", "1000", }; 958 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]); 959 } 960 /* Millisecond latency */ 961 tmp = json_create_object(); 962 json_object_add_value_object(root, "latency_ms", tmp); 963 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) { 964 const char *ranges[] = { "2", "4", "10", "20", "50", "100", 965 "250", "500", "750", "1000", "2000", 966 ">=2000", }; 967 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]); 968 } 969 970 /* Additional output if continue_on_error set - default off*/ 971 if (ts->continue_on_error) { 972 json_object_add_value_int(root, "total_err", ts->total_err_count); 973 json_object_add_value_int(root, "total_err", ts->first_error); 974 } 975 976 /* Additional output if description is set */ 977 if (strlen(ts->description)) 978 json_object_add_value_string(root, "desc", ts->description); 979 980 return root; 981} 982 983static void show_thread_status_terse(struct thread_stat *ts, 984 struct group_run_stats *rs) 985{ 986 if (terse_version == 2) 987 show_thread_status_terse_v2(ts, rs); 988 else if (terse_version == 3 || terse_version == 4) 989 show_thread_status_terse_v3_v4(ts, rs, terse_version); 990 else 991 log_err("fio: bad terse version!? %d\n", terse_version); 992} 993 994static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr) 995{ 996 double mean, S; 997 998 if (src->samples == 0) 999 return; 1000 1001 dst->min_val = min(dst->min_val, src->min_val); 1002 dst->max_val = max(dst->max_val, src->max_val); 1003 1004 /* 1005 * Compute new mean and S after the merge 1006 * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance 1007 * #Parallel_algorithm> 1008 */ 1009 if (nr == 1) { 1010 mean = src->mean.u.f; 1011 S = src->S.u.f; 1012 } else { 1013 double delta = src->mean.u.f - dst->mean.u.f; 1014 1015 mean = ((src->mean.u.f * src->samples) + 1016 (dst->mean.u.f * dst->samples)) / 1017 (dst->samples + src->samples); 1018 1019 S = src->S.u.f + dst->S.u.f + pow(delta, 2.0) * 1020 (dst->samples * src->samples) / 1021 (dst->samples + src->samples); 1022 } 1023 1024 dst->samples += src->samples; 1025 dst->mean.u.f = mean; 1026 dst->S.u.f = S; 1027} 1028 1029void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src) 1030{ 1031 int i; 1032 1033 for (i = 0; i < DDIR_RWDIR_CNT; i++) { 1034 if (dst->max_run[i] < src->max_run[i]) 1035 dst->max_run[i] = src->max_run[i]; 1036 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i]) 1037 dst->min_run[i] = src->min_run[i]; 1038 if (dst->max_bw[i] < src->max_bw[i]) 1039 dst->max_bw[i] = src->max_bw[i]; 1040 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i]) 1041 dst->min_bw[i] = src->min_bw[i]; 1042 1043 dst->io_kb[i] += src->io_kb[i]; 1044 dst->agg[i] += src->agg[i]; 1045 } 1046 1047} 1048 1049void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr) 1050{ 1051 int l, k; 1052 1053 for (l = 0; l < DDIR_RWDIR_CNT; l++) { 1054 if (!dst->unified_rw_rep) { 1055 sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr); 1056 sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr); 1057 sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr); 1058 sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr); 1059 1060 dst->io_bytes[l] += src->io_bytes[l]; 1061 1062 if (dst->runtime[l] < src->runtime[l]) 1063 dst->runtime[l] = src->runtime[l]; 1064 } else { 1065 sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr); 1066 sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr); 1067 sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr); 1068 sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr); 1069 1070 dst->io_bytes[0] += src->io_bytes[l]; 1071 1072 if (dst->runtime[0] < src->runtime[l]) 1073 dst->runtime[0] = src->runtime[l]; 1074 } 1075 } 1076 1077 dst->usr_time += src->usr_time; 1078 dst->sys_time += src->sys_time; 1079 dst->ctx += src->ctx; 1080 dst->majf += src->majf; 1081 dst->minf += src->minf; 1082 1083 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1084 dst->io_u_map[k] += src->io_u_map[k]; 1085 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1086 dst->io_u_submit[k] += src->io_u_submit[k]; 1087 for (k = 0; k < FIO_IO_U_MAP_NR; k++) 1088 dst->io_u_complete[k] += src->io_u_complete[k]; 1089 for (k = 0; k < FIO_IO_U_LAT_U_NR; k++) 1090 dst->io_u_lat_u[k] += src->io_u_lat_u[k]; 1091 for (k = 0; k < FIO_IO_U_LAT_M_NR; k++) 1092 dst->io_u_lat_m[k] += src->io_u_lat_m[k]; 1093 1094 for (k = 0; k < DDIR_RWDIR_CNT; k++) { 1095 if (!dst->unified_rw_rep) { 1096 dst->total_io_u[k] += src->total_io_u[k]; 1097 dst->short_io_u[k] += src->short_io_u[k]; 1098 } else { 1099 dst->total_io_u[0] += src->total_io_u[k]; 1100 dst->short_io_u[0] += src->short_io_u[k]; 1101 } 1102 } 1103 1104 for (k = 0; k < DDIR_RWDIR_CNT; k++) { 1105 int m; 1106 1107 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) { 1108 if (!dst->unified_rw_rep) 1109 dst->io_u_plat[k][m] += src->io_u_plat[k][m]; 1110 else 1111 dst->io_u_plat[0][m] += src->io_u_plat[k][m]; 1112 } 1113 } 1114 1115 dst->total_run_time += src->total_run_time; 1116 dst->total_submit += src->total_submit; 1117 dst->total_complete += src->total_complete; 1118} 1119 1120void init_group_run_stat(struct group_run_stats *gs) 1121{ 1122 int i; 1123 memset(gs, 0, sizeof(*gs)); 1124 1125 for (i = 0; i < DDIR_RWDIR_CNT; i++) 1126 gs->min_bw[i] = gs->min_run[i] = ~0UL; 1127} 1128 1129void init_thread_stat(struct thread_stat *ts) 1130{ 1131 int j; 1132 1133 memset(ts, 0, sizeof(*ts)); 1134 1135 for (j = 0; j < DDIR_RWDIR_CNT; j++) { 1136 ts->lat_stat[j].min_val = -1UL; 1137 ts->clat_stat[j].min_val = -1UL; 1138 ts->slat_stat[j].min_val = -1UL; 1139 ts->bw_stat[j].min_val = -1UL; 1140 } 1141 ts->groupid = -1; 1142} 1143 1144void show_run_stats(void) 1145{ 1146 struct group_run_stats *runstats, *rs; 1147 struct thread_data *td; 1148 struct thread_stat *threadstats, *ts; 1149 int i, j, nr_ts, last_ts, idx; 1150 int kb_base_warned = 0; 1151 int unit_base_warned = 0; 1152 struct json_object *root = NULL; 1153 struct json_array *array = NULL; 1154 1155 runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1)); 1156 1157 for (i = 0; i < groupid + 1; i++) 1158 init_group_run_stat(&runstats[i]); 1159 1160 /* 1161 * find out how many threads stats we need. if group reporting isn't 1162 * enabled, it's one-per-td. 1163 */ 1164 nr_ts = 0; 1165 last_ts = -1; 1166 for_each_td(td, i) { 1167 if (!td->o.group_reporting) { 1168 nr_ts++; 1169 continue; 1170 } 1171 if (last_ts == td->groupid) 1172 continue; 1173 1174 last_ts = td->groupid; 1175 nr_ts++; 1176 } 1177 1178 threadstats = malloc(nr_ts * sizeof(struct thread_stat)); 1179 1180 for (i = 0; i < nr_ts; i++) 1181 init_thread_stat(&threadstats[i]); 1182 1183 j = 0; 1184 last_ts = -1; 1185 idx = 0; 1186 for_each_td(td, i) { 1187 if (idx && (!td->o.group_reporting || 1188 (td->o.group_reporting && last_ts != td->groupid))) { 1189 idx = 0; 1190 j++; 1191 } 1192 1193 last_ts = td->groupid; 1194 1195 ts = &threadstats[j]; 1196 1197 ts->clat_percentiles = td->o.clat_percentiles; 1198 ts->percentile_precision = td->o.percentile_precision; 1199 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list)); 1200 1201 idx++; 1202 ts->members++; 1203 1204 if (ts->groupid == -1) { 1205 /* 1206 * These are per-group shared already 1207 */ 1208 strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE); 1209 if (td->o.description) 1210 strncpy(ts->description, td->o.description, 1211 FIO_JOBNAME_SIZE); 1212 else 1213 memset(ts->description, 0, FIO_JOBNAME_SIZE); 1214 1215 /* 1216 * If multiple entries in this group, this is 1217 * the first member. 1218 */ 1219 ts->thread_number = td->thread_number; 1220 ts->groupid = td->groupid; 1221 1222 /* 1223 * first pid in group, not very useful... 1224 */ 1225 ts->pid = td->pid; 1226 1227 ts->kb_base = td->o.kb_base; 1228 ts->unit_base = td->o.unit_base; 1229 ts->unified_rw_rep = td->o.unified_rw_rep; 1230 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) { 1231 log_info("fio: kb_base differs for jobs in group, using" 1232 " %u as the base\n", ts->kb_base); 1233 kb_base_warned = 1; 1234 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) { 1235 log_info("fio: unit_base differs for jobs in group, using" 1236 " %u as the base\n", ts->unit_base); 1237 unit_base_warned = 1; 1238 } 1239 1240 ts->continue_on_error = td->o.continue_on_error; 1241 ts->total_err_count += td->total_err_count; 1242 ts->first_error = td->first_error; 1243 if (!ts->error) { 1244 if (!td->error && td->o.continue_on_error && 1245 td->first_error) { 1246 ts->error = td->first_error; 1247 strcpy(ts->verror, td->verror); 1248 } else if (td->error) { 1249 ts->error = td->error; 1250 strcpy(ts->verror, td->verror); 1251 } 1252 } 1253 1254 sum_thread_stats(ts, &td->ts, idx); 1255 } 1256 1257 for (i = 0; i < nr_ts; i++) { 1258 unsigned long long bw; 1259 1260 ts = &threadstats[i]; 1261 rs = &runstats[ts->groupid]; 1262 rs->kb_base = ts->kb_base; 1263 rs->unit_base = ts->unit_base; 1264 rs->unified_rw_rep += ts->unified_rw_rep; 1265 1266 for (j = 0; j < DDIR_RWDIR_CNT; j++) { 1267 if (!ts->runtime[j]) 1268 continue; 1269 if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j]) 1270 rs->min_run[j] = ts->runtime[j]; 1271 if (ts->runtime[j] > rs->max_run[j]) 1272 rs->max_run[j] = ts->runtime[j]; 1273 1274 bw = 0; 1275 if (ts->runtime[j]) { 1276 unsigned long runt = ts->runtime[j]; 1277 unsigned long long kb; 1278 1279 kb = ts->io_bytes[j] / rs->kb_base; 1280 bw = kb * 1000 / runt; 1281 } 1282 if (bw < rs->min_bw[j]) 1283 rs->min_bw[j] = bw; 1284 if (bw > rs->max_bw[j]) 1285 rs->max_bw[j] = bw; 1286 1287 rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base; 1288 } 1289 } 1290 1291 for (i = 0; i < groupid + 1; i++) { 1292 int ddir; 1293 1294 rs = &runstats[i]; 1295 1296 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) { 1297 if (rs->max_run[ddir]) 1298 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) / 1299 rs->max_run[ddir]; 1300 } 1301 } 1302 1303 /* 1304 * don't overwrite last signal output 1305 */ 1306 if (output_format == FIO_OUTPUT_NORMAL) 1307 log_info("\n"); 1308 else if (output_format == FIO_OUTPUT_JSON) { 1309 root = json_create_object(); 1310 json_object_add_value_string(root, "fio version", fio_version_string); 1311 array = json_create_array(); 1312 json_object_add_value_array(root, "jobs", array); 1313 } 1314 1315 for (i = 0; i < nr_ts; i++) { 1316 ts = &threadstats[i]; 1317 rs = &runstats[ts->groupid]; 1318 1319 if (is_backend) 1320 fio_server_send_ts(ts, rs); 1321 else if (output_format == FIO_OUTPUT_TERSE) 1322 show_thread_status_terse(ts, rs); 1323 else if (output_format == FIO_OUTPUT_JSON) { 1324 struct json_object *tmp = show_thread_status_json(ts, rs); 1325 json_array_add_value_object(array, tmp); 1326 } else 1327 show_thread_status(ts, rs); 1328 } 1329 if (output_format == FIO_OUTPUT_JSON) { 1330 /* disk util stats, if any */ 1331 show_disk_util(1, root); 1332 1333 show_idle_prof_stats(FIO_OUTPUT_JSON, root); 1334 1335 json_print_object(root); 1336 log_info("\n"); 1337 json_free_object(root); 1338 } 1339 1340 for (i = 0; i < groupid + 1; i++) { 1341 rs = &runstats[i]; 1342 1343 rs->groupid = i; 1344 if (is_backend) 1345 fio_server_send_gs(rs); 1346 else if (output_format == FIO_OUTPUT_NORMAL) 1347 show_group_stats(rs); 1348 } 1349 1350 if (is_backend) 1351 fio_server_send_du(); 1352 else if (output_format == FIO_OUTPUT_NORMAL) { 1353 show_disk_util(0, NULL); 1354 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL); 1355 } 1356 1357 log_info_flush(); 1358 free(runstats); 1359 free(threadstats); 1360} 1361 1362static void *__show_running_run_stats(void fio_unused *arg) 1363{ 1364 struct thread_data *td; 1365 unsigned long long *rt; 1366 struct timeval tv; 1367 int i; 1368 1369 rt = malloc(thread_number * sizeof(unsigned long long)); 1370 fio_gettime(&tv, NULL); 1371 1372 for_each_td(td, i) { 1373 rt[i] = mtime_since(&td->start, &tv); 1374 if (td_read(td) && td->io_bytes[DDIR_READ]) 1375 td->ts.runtime[DDIR_READ] += rt[i]; 1376 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1377 td->ts.runtime[DDIR_WRITE] += rt[i]; 1378 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1379 td->ts.runtime[DDIR_TRIM] += rt[i]; 1380 1381 td->update_rusage = 1; 1382 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; 1383 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; 1384 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; 1385 td->ts.total_run_time = mtime_since(&td->epoch, &tv); 1386 } 1387 1388 for_each_td(td, i) { 1389 if (td->rusage_sem) { 1390 td->update_rusage = 1; 1391 fio_mutex_down(td->rusage_sem); 1392 } 1393 td->update_rusage = 0; 1394 } 1395 1396 show_run_stats(); 1397 1398 for_each_td(td, i) { 1399 if (td_read(td) && td->io_bytes[DDIR_READ]) 1400 td->ts.runtime[DDIR_READ] -= rt[i]; 1401 if (td_write(td) && td->io_bytes[DDIR_WRITE]) 1402 td->ts.runtime[DDIR_WRITE] -= rt[i]; 1403 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) 1404 td->ts.runtime[DDIR_TRIM] -= rt[i]; 1405 } 1406 1407 free(rt); 1408 return NULL; 1409} 1410 1411/* 1412 * Called from signal handler. It _should_ be safe to just run this inline 1413 * in the sig handler, but we should be disturbing the system less by just 1414 * creating a thread to do it. 1415 */ 1416void show_running_run_stats(void) 1417{ 1418 pthread_t thread; 1419 1420 pthread_create(&thread, NULL, __show_running_run_stats, NULL); 1421 pthread_detach(thread); 1422} 1423 1424static int status_interval_init; 1425static struct timeval status_time; 1426 1427#define FIO_STATUS_FILE "/tmp/fio-dump-status" 1428 1429static int check_status_file(void) 1430{ 1431 struct stat sb; 1432 1433 if (stat(FIO_STATUS_FILE, &sb)) 1434 return 0; 1435 1436 unlink(FIO_STATUS_FILE); 1437 return 1; 1438} 1439 1440void check_for_running_stats(void) 1441{ 1442 if (status_interval) { 1443 if (!status_interval_init) { 1444 fio_gettime(&status_time, NULL); 1445 status_interval_init = 1; 1446 } else if (mtime_since_now(&status_time) >= status_interval) { 1447 show_running_run_stats(); 1448 fio_gettime(&status_time, NULL); 1449 return; 1450 } 1451 } 1452 if (check_status_file()) { 1453 show_running_run_stats(); 1454 return; 1455 } 1456} 1457 1458static inline void add_stat_sample(struct io_stat *is, unsigned long data) 1459{ 1460 double val = data; 1461 double delta; 1462 1463 if (data > is->max_val) 1464 is->max_val = data; 1465 if (data < is->min_val) 1466 is->min_val = data; 1467 1468 delta = val - is->mean.u.f; 1469 if (delta) { 1470 is->mean.u.f += delta / (is->samples + 1.0); 1471 is->S.u.f += delta * (val - is->mean.u.f); 1472 } 1473 1474 is->samples++; 1475} 1476 1477static void __add_log_sample(struct io_log *iolog, unsigned long val, 1478 enum fio_ddir ddir, unsigned int bs, 1479 unsigned long t) 1480{ 1481 const int nr_samples = iolog->nr_samples; 1482 1483 if (!iolog->nr_samples) 1484 iolog->avg_last = t; 1485 1486 if (iolog->nr_samples == iolog->max_samples) { 1487 int new_size = sizeof(struct io_sample) * iolog->max_samples*2; 1488 1489 iolog->log = realloc(iolog->log, new_size); 1490 iolog->max_samples <<= 1; 1491 } 1492 1493 iolog->log[nr_samples].val = val; 1494 iolog->log[nr_samples].time = t; 1495 iolog->log[nr_samples].ddir = ddir; 1496 iolog->log[nr_samples].bs = bs; 1497 iolog->nr_samples++; 1498} 1499 1500static inline void reset_io_stat(struct io_stat *ios) 1501{ 1502 ios->max_val = ios->min_val = ios->samples = 0; 1503 ios->mean.u.f = ios->S.u.f = 0; 1504} 1505 1506static void add_log_sample(struct thread_data *td, struct io_log *iolog, 1507 unsigned long val, enum fio_ddir ddir, 1508 unsigned int bs) 1509{ 1510 unsigned long elapsed, this_window; 1511 1512 if (!ddir_rw(ddir)) 1513 return; 1514 1515 elapsed = mtime_since_now(&td->epoch); 1516 1517 /* 1518 * If no time averaging, just add the log sample. 1519 */ 1520 if (!iolog->avg_msec) { 1521 __add_log_sample(iolog, val, ddir, bs, elapsed); 1522 return; 1523 } 1524 1525 /* 1526 * Add the sample. If the time period has passed, then 1527 * add that entry to the log and clear. 1528 */ 1529 add_stat_sample(&iolog->avg_window[ddir], val); 1530 1531 /* 1532 * If period hasn't passed, adding the above sample is all we 1533 * need to do. 1534 */ 1535 this_window = elapsed - iolog->avg_last; 1536 if (this_window < iolog->avg_msec) 1537 return; 1538 1539 /* 1540 * Note an entry in the log. Use the mean from the logged samples, 1541 * making sure to properly round up. Only write a log entry if we 1542 * had actual samples done. 1543 */ 1544 if (iolog->avg_window[DDIR_READ].samples) { 1545 unsigned long mr; 1546 1547 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50; 1548 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed); 1549 } 1550 if (iolog->avg_window[DDIR_WRITE].samples) { 1551 unsigned long mw; 1552 1553 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50; 1554 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed); 1555 } 1556 if (iolog->avg_window[DDIR_TRIM].samples) { 1557 unsigned long mw; 1558 1559 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50; 1560 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed); 1561 } 1562 1563 1564 reset_io_stat(&iolog->avg_window[DDIR_READ]); 1565 reset_io_stat(&iolog->avg_window[DDIR_WRITE]); 1566 reset_io_stat(&iolog->avg_window[DDIR_TRIM]); 1567 iolog->avg_last = elapsed; 1568} 1569 1570void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs) 1571{ 1572 struct io_log *iolog; 1573 1574 if (!ddir_rw(ddir)) 1575 return; 1576 1577 iolog = agg_io_log[ddir]; 1578 __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis()); 1579} 1580 1581static void add_clat_percentile_sample(struct thread_stat *ts, 1582 unsigned long usec, enum fio_ddir ddir) 1583{ 1584 unsigned int idx = plat_val_to_idx(usec); 1585 assert(idx < FIO_IO_U_PLAT_NR); 1586 1587 ts->io_u_plat[ddir][idx]++; 1588} 1589 1590void add_clat_sample(struct thread_data *td, enum fio_ddir ddir, 1591 unsigned long usec, unsigned int bs) 1592{ 1593 struct thread_stat *ts = &td->ts; 1594 1595 if (!ddir_rw(ddir)) 1596 return; 1597 1598 add_stat_sample(&ts->clat_stat[ddir], usec); 1599 1600 if (td->clat_log) 1601 add_log_sample(td, td->clat_log, usec, ddir, bs); 1602 1603 if (ts->clat_percentiles) 1604 add_clat_percentile_sample(ts, usec, ddir); 1605} 1606 1607void add_slat_sample(struct thread_data *td, enum fio_ddir ddir, 1608 unsigned long usec, unsigned int bs) 1609{ 1610 struct thread_stat *ts = &td->ts; 1611 1612 if (!ddir_rw(ddir)) 1613 return; 1614 1615 add_stat_sample(&ts->slat_stat[ddir], usec); 1616 1617 if (td->slat_log) 1618 add_log_sample(td, td->slat_log, usec, ddir, bs); 1619} 1620 1621void add_lat_sample(struct thread_data *td, enum fio_ddir ddir, 1622 unsigned long usec, unsigned int bs) 1623{ 1624 struct thread_stat *ts = &td->ts; 1625 1626 if (!ddir_rw(ddir)) 1627 return; 1628 1629 add_stat_sample(&ts->lat_stat[ddir], usec); 1630 1631 if (td->lat_log) 1632 add_log_sample(td, td->lat_log, usec, ddir, bs); 1633} 1634 1635void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs, 1636 struct timeval *t) 1637{ 1638 struct thread_stat *ts = &td->ts; 1639 unsigned long spent, rate; 1640 1641 if (!ddir_rw(ddir)) 1642 return; 1643 1644 spent = mtime_since(&td->bw_sample_time, t); 1645 if (spent < td->o.bw_avg_time) 1646 return; 1647 1648 /* 1649 * Compute both read and write rates for the interval. 1650 */ 1651 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1652 uint64_t delta; 1653 1654 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]; 1655 if (!delta) 1656 continue; /* No entries for interval */ 1657 1658 rate = delta * 1000 / spent / 1024; 1659 add_stat_sample(&ts->bw_stat[ddir], rate); 1660 1661 if (td->bw_log) 1662 add_log_sample(td, td->bw_log, rate, ddir, bs); 1663 1664 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir]; 1665 } 1666 1667 fio_gettime(&td->bw_sample_time, NULL); 1668} 1669 1670void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, 1671 struct timeval *t) 1672{ 1673 struct thread_stat *ts = &td->ts; 1674 unsigned long spent, iops; 1675 1676 if (!ddir_rw(ddir)) 1677 return; 1678 1679 spent = mtime_since(&td->iops_sample_time, t); 1680 if (spent < td->o.iops_avg_time) 1681 return; 1682 1683 /* 1684 * Compute both read and write rates for the interval. 1685 */ 1686 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) { 1687 uint64_t delta; 1688 1689 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir]; 1690 if (!delta) 1691 continue; /* No entries for interval */ 1692 1693 iops = (delta * 1000) / spent; 1694 add_stat_sample(&ts->iops_stat[ddir], iops); 1695 1696 if (td->iops_log) 1697 add_log_sample(td, td->iops_log, iops, ddir, 0); 1698 1699 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir]; 1700 } 1701 1702 fio_gettime(&td->iops_sample_time, NULL); 1703} 1704