stat.c

   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 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 struct fio_mutex *stat_mutex;
  22
  23 void clear_rusage_stat(struct thread_data *td)
  24 {
  25         struct thread_stat *ts = &td->ts;
  26
  27         fio_getrusage(&td->ru_start);
  28         ts->usr_time = ts->sys_time = 0;
  29         ts->ctx = 0;
  30         ts->minf = ts->majf = 0;
  31 }
  32
  33 void update_rusage_stat(struct thread_data *td)
  34 {
  35         struct thread_stat *ts = &td->ts;
  36
  37         fio_getrusage(&td->ru_end);
  38         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  39                                         &td->ru_end.ru_utime);
  40         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  41                                         &td->ru_end.ru_stime);
  42         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  43                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  44         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  45         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  46
  47         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  48 }
  49
  50 /*
  51  * Given a latency, return the index of the corresponding bucket in
  52  * the structure tracking percentiles.
  53  *
  54  * (1) find the group (and error bits) that the value (latency)
  55  * belongs to by looking at its MSB. (2) find the bucket number in the
  56  * group by looking at the index bits.
  57  *
  58  */
  59 static unsigned int plat_val_to_idx(unsigned int val)
  60 {
  61         unsigned int msb, error_bits, base, offset, idx;
  62
  63         /* Find MSB starting from bit 0 */
  64         if (val == 0)
  65                 msb = 0;
  66         else
  67                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  68
  69         /*
  70          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  71          * all bits of the sample as index
  72          */
  73         if (msb <= FIO_IO_U_PLAT_BITS)
  74                 return val;
  75
  76         /* Compute the number of error bits to discard*/
  77         error_bits = msb - FIO_IO_U_PLAT_BITS;
  78
  79         /* Compute the number of buckets before the group */
  80         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  81
  82         /*
  83          * Discard the error bits and apply the mask to find the
  84          * index for the buckets in the group
  85          */
  86         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  87
  88         /* Make sure the index does not exceed (array size - 1) */
  89         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  90                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  91
  92         return idx;
  93 }
  94
  95 /*
  96  * Convert the given index of the bucket array to the value
  97  * represented by the bucket
  98  */
  99 static unsigned int plat_idx_to_val(unsigned int idx)
 100 {
 101         unsigned int error_bits, k, base;
 102
 103         assert(idx < FIO_IO_U_PLAT_NR);
 104
 105         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 106          * all bits of the sample as index */
 107         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 108                 return idx;
 109
 110         /* Find the group and compute the minimum value of that group */
 111         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 112         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 113
 114         /* Find its bucket number of the group */
 115         k = idx % FIO_IO_U_PLAT_VAL;
 116
 117         /* Return the mean of the range of the bucket */
 118         return base + ((k + 0.5) * (1 << error_bits));
 119 }
 120
 121 static int double_cmp(const void *a, const void *b)
 122 {
 123         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 124         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 125         int cmp = 0;
 126
 127         if (fa.u.f > fb.u.f)
 128                 cmp = 1;
 129         else if (fa.u.f < fb.u.f)
 130                 cmp = -1;
 131
 132         return cmp;
 133 }
 134
 135 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 136                                    fio_fp64_t *plist, unsigned int **output,
 137                                    unsigned int *maxv, unsigned int *minv)
 138 {
 139         unsigned long sum = 0;
 140         unsigned int len, i, j = 0;
 141         unsigned int oval_len = 0;
 142         unsigned int *ovals = NULL;
 143         int is_last;
 144
 145         *minv = -1U;
 146         *maxv = 0;
 147
 148         len = 0;
 149         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 150                 len++;
 151
 152         if (!len)
 153                 return 0;
 154
 155         /*
 156          * Sort the percentile list. Note that it may already be sorted if
 157          * we are using the default values, but since it's a short list this
 158          * isn't a worry. Also note that this does not work for NaN values.
 159          */
 160         if (len > 1)
 161                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 162
 163         /*
 164          * Calculate bucket values, note down max and min values
 165          */
 166         is_last = 0;
 167         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 168                 sum += io_u_plat[i];
 169                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 170                         assert(plist[j].u.f <= 100.0);
 171
 172                         if (j == oval_len) {
 173                                 oval_len += 100;
 174                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 175                         }
 176
 177                         ovals[j] = plat_idx_to_val(i);
 178                         if (ovals[j] < *minv)
 179                                 *minv = ovals[j];
 180                         if (ovals[j] > *maxv)
 181                                 *maxv = ovals[j];
 182
 183                         is_last = (j == len - 1);
 184                         if (is_last)
 185                                 break;
 186
 187                         j++;
 188                 }
 189         }
 190
 191         *output = ovals;
 192         return len;
 193 }
 194
 195 /*
 196  * Find and display the p-th percentile of clat
 197  */
 198 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 199                                   fio_fp64_t *plist, unsigned int precision,
 200                                   struct buf_output *out)
 201 {
 202         unsigned int len, j = 0, minv, maxv;
 203         unsigned int *ovals;
 204         int is_last, per_line, scale_down;
 205         char fmt[32];
 206
 207         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 208         if (!len)
 209                 goto out;
 210
 211         /*
 212          * We default to usecs, but if the value range is such that we
 213          * should scale down to msecs, do that.
 214          */
 215         if (minv > 2000 && maxv > 99999) {
 216                 scale_down = 1;
 217                 log_buf(out, "    clat percentiles (msec):\n     |");
 218         } else {
 219                 scale_down = 0;
 220                 log_buf(out, "    clat percentiles (usec):\n     |");
 221         }
 222
 223         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 224         per_line = (80 - 7) / (precision + 14);
 225
 226         for (j = 0; j < len; j++) {
 227                 char fbuf[16], *ptr = fbuf;
 228
 229                 /* for formatting */
 230                 if (j != 0 && (j % per_line) == 0)
 231                         log_buf(out, "     |");
 232
 233                 /* end of the list */
 234                 is_last = (j == len - 1);
 235
 236                 if (plist[j].u.f < 10.0)
 237                         ptr += sprintf(fbuf, " ");
 238
 239                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 240
 241                 if (scale_down)
 242                         ovals[j] = (ovals[j] + 999) / 1000;
 243
 244                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 245
 246                 if (is_last)
 247                         break;
 248
 249                 if ((j % per_line) == per_line - 1)     /* for formatting */
 250                         log_buf(out, "\n");
 251         }
 252
 253 out:
 254         if (ovals)
 255                 free(ovals);
 256 }
 257
 258 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 259              double *mean, double *dev)
 260 {
 261         double n = (double) is->samples;
 262
 263         if (n == 0)
 264                 return 0;
 265
 266         *min = is->min_val;
 267         *max = is->max_val;
 268         *mean = is->mean.u.f;
 269
 270         if (n > 1.0)
 271                 *dev = sqrt(is->S.u.f / (n - 1.0));
 272         else
 273                 *dev = 0;
 274
 275         return 1;
 276 }
 277
 278 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 279 {
 280         char *p1, *p2, *p3, *p4;
 281         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 282         int i;
 283
 284         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 285
 286         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 287                 const int i2p = is_power_of_2(rs->kb_base);
 288
 289                 if (!rs->max_run[i])
 290                         continue;
 291
 292                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 293                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 294                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 295                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 296
 297                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 298                          " mint=%llumsec, maxt=%llumsec\n",
 299                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 300                                 p1, p2, p3, p4,
 301                                 (unsigned long long) rs->min_run[i],
 302                                 (unsigned long long) rs->max_run[i]);
 303
 304                 free(p1);
 305                 free(p2);
 306                 free(p3);
 307                 free(p4);
 308         }
 309 }
 310
 311 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 312 {
 313         int i;
 314
 315         /*
 316          * Do depth distribution calculations
 317          */
 318         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 319                 if (total) {
 320                         io_u_dist[i] = (double) map[i] / (double) total;
 321                         io_u_dist[i] *= 100.0;
 322                         if (io_u_dist[i] < 0.1 && map[i])
 323                                 io_u_dist[i] = 0.1;
 324                 } else
 325                         io_u_dist[i] = 0.0;
 326         }
 327 }
 328
 329 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 330                           unsigned int *src, int nr)
 331 {
 332         unsigned long total = ddir_rw_sum(ts->total_io_u);
 333         int i;
 334
 335         /*
 336          * Do latency distribution calculations
 337          */
 338         for (i = 0; i < nr; i++) {
 339                 if (total) {
 340                         dst[i] = (double) src[i] / (double) total;
 341                         dst[i] *= 100.0;
 342                         if (dst[i] < 0.01 && src[i])
 343                                 dst[i] = 0.01;
 344                 } else
 345                         dst[i] = 0.0;
 346         }
 347 }
 348
 349 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 350 {
 351         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 352 }
 353
 354 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 355 {
 356         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 357 }
 358
 359 static void display_lat(const char *name, unsigned long min, unsigned long max,
 360                         double mean, double dev, struct buf_output *out)
 361 {
 362         const char *base = "(usec)";
 363         char *minp, *maxp;
 364
 365         if (!usec_to_msec(&min, &max, &mean, &dev))
 366                 base = "(msec)";
 367
 368         minp = num2str(min, 6, 1, 0, 0);
 369         maxp = num2str(max, 6, 1, 0, 0);
 370
 371         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 372                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 373
 374         free(minp);
 375         free(maxp);
 376 }
 377
 378 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 379                              int ddir, struct buf_output *out)
 380 {
 381         const char *str[] = { "read ", "write", "trim" };
 382         unsigned long min, max, runt;
 383         unsigned long long bw, iops;
 384         double mean, dev;
 385         char *io_p, *bw_p, *iops_p;
 386         int i2p;
 387
 388         assert(ddir_rw(ddir));
 389
 390         if (!ts->runtime[ddir])
 391                 return;
 392
 393         i2p = is_power_of_2(rs->kb_base);
 394         runt = ts->runtime[ddir];
 395
 396         bw = (1000 * ts->io_bytes[ddir]) / runt;
 397         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 398         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 399
 400         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 401         iops_p = num2str(iops, 6, 1, 0, 0);
 402
 403         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 404                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 405                                 io_p, bw_p, iops_p,
 406                                 (unsigned long long) ts->runtime[ddir]);
 407
 408         free(io_p);
 409         free(bw_p);
 410         free(iops_p);
 411
 412         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 413                 display_lat("slat", min, max, mean, dev, out);
 414         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 415                 display_lat("clat", min, max, mean, dev, out);
 416         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 417                 display_lat(" lat", min, max, mean, dev, out);
 418
 419         if (ts->clat_percentiles) {
 420                 show_clat_percentiles(ts->io_u_plat[ddir],
 421                                         ts->clat_stat[ddir].samples,
 422                                         ts->percentile_list,
 423                                         ts->percentile_precision, out);
 424         }
 425         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 426                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 427                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 428
 429                 if (rs->unit_base == 1) {
 430                         min *= 8.0;
 431                         max *= 8.0;
 432                         mean *= 8.0;
 433                         dev *= 8.0;
 434                 }
 435
 436                 if (rs->agg[ddir]) {
 437                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 438                         if (p_of_agg > 100.0)
 439                                 p_of_agg = 100.0;
 440                 }
 441
 442                 if (mean > fkb_base * fkb_base) {
 443                         min /= fkb_base;
 444                         max /= fkb_base;
 445                         mean /= fkb_base;
 446                         dev /= fkb_base;
 447                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 448                 }
 449
 450                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 451                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 452                                                         p_of_agg, mean, dev);
 453         }
 454 }
 455
 456 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 457                     const char *msg, struct buf_output *out)
 458 {
 459         int new_line = 1, i, line = 0, shown = 0;
 460
 461         for (i = 0; i < nr; i++) {
 462                 if (io_u_lat[i] <= 0.0)
 463                         continue;
 464                 shown = 1;
 465                 if (new_line) {
 466                         if (line)
 467                                 log_buf(out, "\n");
 468                         log_buf(out, "    lat (%s) : ", msg);
 469                         new_line = 0;
 470                         line = 0;
 471                 }
 472                 if (line)
 473                         log_buf(out, ", ");
 474                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 475                 line++;
 476                 if (line == 5)
 477                         new_line = 1;
 478         }
 479
 480         if (shown)
 481                 log_buf(out, "\n");
 482
 483         return shown;
 484 }
 485
 486 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 487 {
 488         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 489                                  "250=", "500=", "750=", "1000=", };
 490
 491         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 492 }
 493
 494 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 495 {
 496         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 497                                  "250=", "500=", "750=", "1000=", "2000=",
 498                                  ">=2000=", };
 499
 500         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 501 }
 502
 503 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 504 {
 505         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 506         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 507
 508         stat_calc_lat_u(ts, io_u_lat_u);
 509         stat_calc_lat_m(ts, io_u_lat_m);
 510
 511         show_lat_u(io_u_lat_u, out);
 512         show_lat_m(io_u_lat_m, out);
 513 }
 514
 515 static int block_state_category(int block_state)
 516 {
 517         switch (block_state) {
 518         case BLOCK_STATE_UNINIT:
 519                 return 0;
 520         case BLOCK_STATE_TRIMMED:
 521         case BLOCK_STATE_WRITTEN:
 522                 return 1;
 523         case BLOCK_STATE_WRITE_FAILURE:
 524         case BLOCK_STATE_TRIM_FAILURE:
 525                 return 2;
 526         default:
 527                 /* Silence compile warning on some BSDs and have a return */
 528                 assert(0);
 529                 return -1;
 530         }
 531 }
 532
 533 static int compare_block_infos(const void *bs1, const void *bs2)
 534 {
 535         uint32_t block1 = *(uint32_t *)bs1;
 536         uint32_t block2 = *(uint32_t *)bs2;
 537         int state1 = BLOCK_INFO_STATE(block1);
 538         int state2 = BLOCK_INFO_STATE(block2);
 539         int bscat1 = block_state_category(state1);
 540         int bscat2 = block_state_category(state2);
 541         int cycles1 = BLOCK_INFO_TRIMS(block1);
 542         int cycles2 = BLOCK_INFO_TRIMS(block2);
 543
 544         if (bscat1 < bscat2)
 545                 return -1;
 546         if (bscat1 > bscat2)
 547                 return 1;
 548
 549         if (cycles1 < cycles2)
 550                 return -1;
 551         if (cycles1 > cycles2)
 552                 return 1;
 553
 554         if (state1 < state2)
 555                 return -1;
 556         if (state1 > state2)
 557                 return 1;
 558
 559         assert(block1 == block2);
 560         return 0;
 561 }
 562
 563 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 564                                   fio_fp64_t *plist, unsigned int **percentiles,
 565                                   unsigned int *types)
 566 {
 567         int len = 0;
 568         int i, nr_uninit;
 569
 570         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 571
 572         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 573                 len++;
 574
 575         if (!len)
 576                 return 0;
 577
 578         /*
 579          * Sort the percentile list. Note that it may already be sorted if
 580          * we are using the default values, but since it's a short list this
 581          * isn't a worry. Also note that this does not work for NaN values.
 582          */
 583         if (len > 1)
 584                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 585
 586         nr_uninit = 0;
 587         /* Start only after the uninit entries end */
 588         for (nr_uninit = 0;
 589              nr_uninit < nr_block_infos
 590                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 591              nr_uninit ++)
 592                 ;
 593
 594         if (nr_uninit == nr_block_infos)
 595                 return 0;
 596
 597         *percentiles = calloc(len, sizeof(**percentiles));
 598
 599         for (i = 0; i < len; i++) {
 600                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 601                                 + nr_uninit;
 602                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 603         }
 604
 605         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 606         for (i = 0; i < nr_block_infos; i++)
 607                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 608
 609         return len;
 610 }
 611
 612 static const char *block_state_names[] = {
 613         [BLOCK_STATE_UNINIT] = "unwritten",
 614         [BLOCK_STATE_TRIMMED] = "trimmed",
 615         [BLOCK_STATE_WRITTEN] = "written",
 616         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 617         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 618 };
 619
 620 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 621                              fio_fp64_t *plist, struct buf_output *out)
 622 {
 623         int len, pos, i;
 624         unsigned int *percentiles = NULL;
 625         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 626
 627         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 628                                      &percentiles, block_state_counts);
 629
 630         log_buf(out, "  block lifetime percentiles :\n   |");
 631         pos = 0;
 632         for (i = 0; i < len; i++) {
 633                 uint32_t block_info = percentiles[i];
 634 #define LINE_LENGTH     75
 635                 char str[LINE_LENGTH];
 636                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 637                                      plist[i].u.f, block_info,
 638                                      i == len - 1 ? '\n' : ',');
 639                 assert(strln < LINE_LENGTH);
 640                 if (pos + strln > LINE_LENGTH) {
 641                         pos = 0;
 642                         log_buf(out, "\n   |");
 643                 }
 644                 log_buf(out, "%s", str);
 645                 pos += strln;
 646 #undef LINE_LENGTH
 647         }
 648         if (percentiles)
 649                 free(percentiles);
 650
 651         log_buf(out, "        states               :");
 652         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 653                 log_buf(out, " %s=%u%c",
 654                          block_state_names[i], block_state_counts[i],
 655                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 656 }
 657
 658 static void show_thread_status_normal(struct thread_stat *ts,
 659                                       struct group_run_stats *rs,
 660                                       struct buf_output *out)
 661 {
 662         double usr_cpu, sys_cpu;
 663         unsigned long runtime;
 664         double io_u_dist[FIO_IO_U_MAP_NR];
 665         time_t time_p;
 666         char time_buf[32];
 667
 668         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 669                 return;
 670
 671         time(&time_p);
 672         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 673
 674         if (!ts->error) {
 675                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 676                                         ts->name, ts->groupid, ts->members,
 677                                         ts->error, (int) ts->pid, time_buf);
 678         } else {
 679                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 680                                         ts->name, ts->groupid, ts->members,
 681                                         ts->error, ts->verror, (int) ts->pid,
 682                                         time_buf);
 683         }
 684
 685         if (strlen(ts->description))
 686                 log_buf(out, "  Description  : [%s]\n", ts->description);
 687
 688         if (ts->io_bytes[DDIR_READ])
 689                 show_ddir_status(rs, ts, DDIR_READ, out);
 690         if (ts->io_bytes[DDIR_WRITE])
 691                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 692         if (ts->io_bytes[DDIR_TRIM])
 693                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 694
 695         show_latencies(ts, out);
 696
 697         runtime = ts->total_run_time;
 698         if (runtime) {
 699                 double runt = (double) runtime;
 700
 701                 usr_cpu = (double) ts->usr_time * 100 / runt;
 702                 sys_cpu = (double) ts->sys_time * 100 / runt;
 703         } else {
 704                 usr_cpu = 0;
 705                 sys_cpu = 0;
 706         }
 707
 708         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 709                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 710                         (unsigned long long) ts->ctx,
 711                         (unsigned long long) ts->majf,
 712                         (unsigned long long) ts->minf);
 713
 714         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 715         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 716                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 717                                         io_u_dist[1], io_u_dist[2],
 718                                         io_u_dist[3], io_u_dist[4],
 719                                         io_u_dist[5], io_u_dist[6]);
 720
 721         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 722         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 723                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 724                                         io_u_dist[1], io_u_dist[2],
 725                                         io_u_dist[3], io_u_dist[4],
 726                                         io_u_dist[5], io_u_dist[6]);
 727         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 728         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 729                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 730                                         io_u_dist[1], io_u_dist[2],
 731                                         io_u_dist[3], io_u_dist[4],
 732                                         io_u_dist[5], io_u_dist[6]);
 733         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 734                                  " short=r=%llu/w=%llu/d=%llu,"
 735                                  " drop=r=%llu/w=%llu/d=%llu\n",
 736                                         (unsigned long long) ts->total_io_u[0],
 737                                         (unsigned long long) ts->total_io_u[1],
 738                                         (unsigned long long) ts->total_io_u[2],
 739                                         (unsigned long long) ts->short_io_u[0],
 740                                         (unsigned long long) ts->short_io_u[1],
 741                                         (unsigned long long) ts->short_io_u[2],
 742                                         (unsigned long long) ts->drop_io_u[0],
 743                                         (unsigned long long) ts->drop_io_u[1],
 744                                         (unsigned long long) ts->drop_io_u[2]);
 745         if (ts->continue_on_error) {
 746                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 747                                         (unsigned long long)ts->total_err_count,
 748                                         ts->first_error,
 749                                         strerror(ts->first_error));
 750         }
 751         if (ts->latency_depth) {
 752                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 753                                         (unsigned long long)ts->latency_target,
 754                                         (unsigned long long)ts->latency_window,
 755                                         ts->latency_percentile.u.f,
 756                                         ts->latency_depth);
 757         }
 758
 759         if (ts->nr_block_infos)
 760                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 761                                   ts->percentile_list, out);
 762 }
 763
 764 static void show_ddir_status_terse(struct thread_stat *ts,
 765                                    struct group_run_stats *rs, int ddir,
 766                                    struct buf_output *out)
 767 {
 768         unsigned long min, max;
 769         unsigned long long bw, iops;
 770         unsigned int *ovals = NULL;
 771         double mean, dev;
 772         unsigned int len, minv, maxv;
 773         int i;
 774
 775         assert(ddir_rw(ddir));
 776
 777         iops = bw = 0;
 778         if (ts->runtime[ddir]) {
 779                 uint64_t runt = ts->runtime[ddir];
 780
 781                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 782                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 783         }
 784
 785         log_buf(out, ";%llu;%llu;%llu;%llu",
 786                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 787                                         (unsigned long long) ts->runtime[ddir]);
 788
 789         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 790                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 791         else
 792                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 793
 794         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 795                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 796         else
 797                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 798
 799         if (ts->clat_percentiles) {
 800                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 801                                         ts->clat_stat[ddir].samples,
 802                                         ts->percentile_list, &ovals, &maxv,
 803                                         &minv);
 804         } else
 805                 len = 0;
 806
 807         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 808                 if (i >= len) {
 809                         log_buf(out, ";0%%=0");
 810                         continue;
 811                 }
 812                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 813         }
 814
 815         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 816                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 817         else
 818                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 819
 820         if (ovals)
 821                 free(ovals);
 822
 823         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 824                 double p_of_agg = 100.0;
 825
 826                 if (rs->agg[ddir]) {
 827                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 828                         if (p_of_agg > 100.0)
 829                                 p_of_agg = 100.0;
 830                 }
 831
 832                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 833         } else
 834                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 835 }
 836
 837 static void add_ddir_status_json(struct thread_stat *ts,
 838                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 839 {
 840         unsigned long min, max;
 841         unsigned long long bw;
 842         unsigned int *ovals = NULL;
 843         double mean, dev, iops;
 844         unsigned int len, minv, maxv;
 845         int i;
 846         const char *ddirname[] = {"read", "write", "trim"};
 847         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 848         char buf[120];
 849         double p_of_agg = 100.0;
 850
 851         assert(ddir_rw(ddir));
 852
 853         if (ts->unified_rw_rep && ddir != DDIR_READ)
 854                 return;
 855
 856         dir_object = json_create_object();
 857         json_object_add_value_object(parent,
 858                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 859
 860         bw = 0;
 861         iops = 0.0;
 862         if (ts->runtime[ddir]) {
 863                 uint64_t runt = ts->runtime[ddir];
 864
 865                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 866                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 867         }
 868
 869         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 870         json_object_add_value_int(dir_object, "bw", bw);
 871         json_object_add_value_float(dir_object, "iops", iops);
 872         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 873         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 874         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 875         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 876
 877         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 878                 min = max = 0;
 879                 mean = dev = 0.0;
 880         }
 881         tmp_object = json_create_object();
 882         json_object_add_value_object(dir_object, "slat", tmp_object);
 883         json_object_add_value_int(tmp_object, "min", min);
 884         json_object_add_value_int(tmp_object, "max", max);
 885         json_object_add_value_float(tmp_object, "mean", mean);
 886         json_object_add_value_float(tmp_object, "stddev", dev);
 887
 888         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 889                 min = max = 0;
 890                 mean = dev = 0.0;
 891         }
 892         tmp_object = json_create_object();
 893         json_object_add_value_object(dir_object, "clat", tmp_object);
 894         json_object_add_value_int(tmp_object, "min", min);
 895         json_object_add_value_int(tmp_object, "max", max);
 896         json_object_add_value_float(tmp_object, "mean", mean);
 897         json_object_add_value_float(tmp_object, "stddev", dev);
 898
 899         if (ts->clat_percentiles) {
 900                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 901                                         ts->clat_stat[ddir].samples,
 902                                         ts->percentile_list, &ovals, &maxv,
 903                                         &minv);
 904         } else
 905                 len = 0;
 906
 907         percentile_object = json_create_object();
 908         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 909         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 910                 if (i >= len) {
 911                         json_object_add_value_int(percentile_object, "0.00", 0);
 912                         continue;
 913                 }
 914                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 915                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 916         }
 917
 918         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 919                 clat_bins_object = json_create_object();
 920                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 921                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 922                         snprintf(buf, sizeof(buf), "%d", i);
 923                         json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 924                 }
 925                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
 926                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
 927                 json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
 928         }
 929
 930         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 931                 min = max = 0;
 932                 mean = dev = 0.0;
 933         }
 934         tmp_object = json_create_object();
 935         json_object_add_value_object(dir_object, "lat", tmp_object);
 936         json_object_add_value_int(tmp_object, "min", min);
 937         json_object_add_value_int(tmp_object, "max", max);
 938         json_object_add_value_float(tmp_object, "mean", mean);
 939         json_object_add_value_float(tmp_object, "stddev", dev);
 940         if (ovals)
 941                 free(ovals);
 942
 943         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 944                 if (rs->agg[ddir]) {
 945                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 946                         if (p_of_agg > 100.0)
 947                                 p_of_agg = 100.0;
 948                 }
 949         } else {
 950                 min = max = 0;
 951                 p_of_agg = mean = dev = 0.0;
 952         }
 953         json_object_add_value_int(dir_object, "bw_min", min);
 954         json_object_add_value_int(dir_object, "bw_max", max);
 955         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 956         json_object_add_value_float(dir_object, "bw_mean", mean);
 957         json_object_add_value_float(dir_object, "bw_dev", dev);
 958 }
 959
 960 static void show_thread_status_terse_v2(struct thread_stat *ts,
 961                                         struct group_run_stats *rs,
 962                                         struct buf_output *out)
 963 {
 964         double io_u_dist[FIO_IO_U_MAP_NR];
 965         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 966         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 967         double usr_cpu, sys_cpu;
 968         int i;
 969
 970         /* General Info */
 971         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 972         /* Log Read Status */
 973         show_ddir_status_terse(ts, rs, DDIR_READ, out);
 974         /* Log Write Status */
 975         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
 976         /* Log Trim Status */
 977         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
 978
 979         /* CPU Usage */
 980         if (ts->total_run_time) {
 981                 double runt = (double) ts->total_run_time;
 982
 983                 usr_cpu = (double) ts->usr_time * 100 / runt;
 984                 sys_cpu = (double) ts->sys_time * 100 / runt;
 985         } else {
 986                 usr_cpu = 0;
 987                 sys_cpu = 0;
 988         }
 989
 990         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 991                                                 (unsigned long long) ts->ctx,
 992                                                 (unsigned long long) ts->majf,
 993                                                 (unsigned long long) ts->minf);
 994
 995         /* Calc % distribution of IO depths, usecond, msecond latency */
 996         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 997         stat_calc_lat_u(ts, io_u_lat_u);
 998         stat_calc_lat_m(ts, io_u_lat_m);
 999
1000         /* Only show fixed 7 I/O depth levels*/
1001         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1002                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1003                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1004
1005         /* Microsecond latency */
1006         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1007                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1008         /* Millisecond latency */
1009         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1010                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1011         /* Additional output if continue_on_error set - default off*/
1012         if (ts->continue_on_error)
1013                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1014         log_buf(out, "\n");
1015
1016         /* Additional output if description is set */
1017         if (strlen(ts->description))
1018                 log_buf(out, ";%s", ts->description);
1019
1020         log_buf(out, "\n");
1021 }
1022
1023 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1024                                            struct group_run_stats *rs, int ver,
1025                                            struct buf_output *out)
1026 {
1027         double io_u_dist[FIO_IO_U_MAP_NR];
1028         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1029         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1030         double usr_cpu, sys_cpu;
1031         int i;
1032
1033         /* General Info */
1034         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1035                                         ts->name, ts->groupid, ts->error);
1036         /* Log Read Status */
1037         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1038         /* Log Write Status */
1039         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1040         /* Log Trim Status */
1041         if (ver == 4)
1042                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1043
1044         /* CPU Usage */
1045         if (ts->total_run_time) {
1046                 double runt = (double) ts->total_run_time;
1047
1048                 usr_cpu = (double) ts->usr_time * 100 / runt;
1049                 sys_cpu = (double) ts->sys_time * 100 / runt;
1050         } else {
1051                 usr_cpu = 0;
1052                 sys_cpu = 0;
1053         }
1054
1055         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1056                                                 (unsigned long long) ts->ctx,
1057                                                 (unsigned long long) ts->majf,
1058                                                 (unsigned long long) ts->minf);
1059
1060         /* Calc % distribution of IO depths, usecond, msecond latency */
1061         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1062         stat_calc_lat_u(ts, io_u_lat_u);
1063         stat_calc_lat_m(ts, io_u_lat_m);
1064
1065         /* Only show fixed 7 I/O depth levels*/
1066         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1067                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1068                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1069
1070         /* Microsecond latency */
1071         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1072                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1073         /* Millisecond latency */
1074         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1075                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1076
1077         /* disk util stats, if any */
1078         show_disk_util(1, NULL, out);
1079
1080         /* Additional output if continue_on_error set - default off*/
1081         if (ts->continue_on_error)
1082                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1083
1084         /* Additional output if description is set */
1085         if (strlen(ts->description))
1086                 log_buf(out, ";%s", ts->description);
1087
1088         log_buf(out, "\n");
1089 }
1090
1091 void json_add_job_opts(struct json_object *root, const char *name,
1092                        struct flist_head *opt_list, bool num_jobs)
1093 {
1094         struct json_object *dir_object;
1095         struct flist_head *entry;
1096         struct print_option *p;
1097
1098         if (flist_empty(opt_list))
1099                 return;
1100
1101         dir_object = json_create_object();
1102         json_object_add_value_object(root, name, dir_object);
1103
1104         flist_for_each(entry, opt_list) {
1105                 const char *pos = "";
1106
1107                 p = flist_entry(entry, struct print_option, list);
1108                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1109                         continue;
1110                 if (p->value)
1111                         pos = p->value;
1112                 json_object_add_value_string(dir_object, p->name, pos);
1113         }
1114 }
1115
1116 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1117                                                    struct group_run_stats *rs,
1118                                                    struct flist_head *opt_list)
1119 {
1120         struct json_object *root, *tmp;
1121         struct jobs_eta *je;
1122         double io_u_dist[FIO_IO_U_MAP_NR];
1123         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1124         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1125         double usr_cpu, sys_cpu;
1126         int i;
1127         size_t size;
1128
1129         root = json_create_object();
1130         json_object_add_value_string(root, "jobname", ts->name);
1131         json_object_add_value_int(root, "groupid", ts->groupid);
1132         json_object_add_value_int(root, "error", ts->error);
1133
1134         /* ETA Info */
1135         je = get_jobs_eta(true, &size);
1136         if (je) {
1137                 json_object_add_value_int(root, "eta", je->eta_sec);
1138                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1139         }
1140
1141         if (opt_list)
1142                 json_add_job_opts(root, "job options", opt_list, true);
1143
1144         add_ddir_status_json(ts, rs, DDIR_READ, root);
1145         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1146         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1147
1148         /* CPU Usage */
1149         if (ts->total_run_time) {
1150                 double runt = (double) ts->total_run_time;
1151
1152                 usr_cpu = (double) ts->usr_time * 100 / runt;
1153                 sys_cpu = (double) ts->sys_time * 100 / runt;
1154         } else {
1155                 usr_cpu = 0;
1156                 sys_cpu = 0;
1157         }
1158         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1159         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1160         json_object_add_value_int(root, "ctx", ts->ctx);
1161         json_object_add_value_int(root, "majf", ts->majf);
1162         json_object_add_value_int(root, "minf", ts->minf);
1163
1164
1165         /* Calc % distribution of IO depths, usecond, msecond latency */
1166         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1167         stat_calc_lat_u(ts, io_u_lat_u);
1168         stat_calc_lat_m(ts, io_u_lat_m);
1169
1170         tmp = json_create_object();
1171         json_object_add_value_object(root, "iodepth_level", tmp);
1172         /* Only show fixed 7 I/O depth levels*/
1173         for (i = 0; i < 7; i++) {
1174                 char name[20];
1175                 if (i < 6)
1176                         snprintf(name, 20, "%d", 1 << i);
1177                 else
1178                         snprintf(name, 20, ">=%d", 1 << i);
1179                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1180         }
1181
1182         tmp = json_create_object();
1183         json_object_add_value_object(root, "latency_us", tmp);
1184         /* Microsecond latency */
1185         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1186                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1187                                  "250", "500", "750", "1000", };
1188                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1189         }
1190         /* Millisecond latency */
1191         tmp = json_create_object();
1192         json_object_add_value_object(root, "latency_ms", tmp);
1193         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1194                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1195                                  "250", "500", "750", "1000", "2000",
1196                                  ">=2000", };
1197                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1198         }
1199
1200         /* Additional output if continue_on_error set - default off*/
1201         if (ts->continue_on_error) {
1202                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1203                 json_object_add_value_int(root, "first_error", ts->first_error);
1204         }
1205
1206         if (ts->latency_depth) {
1207                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1208                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1209                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1210                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1211         }
1212
1213         /* Additional output if description is set */
1214         if (strlen(ts->description))
1215                 json_object_add_value_string(root, "desc", ts->description);
1216
1217         if (ts->nr_block_infos) {
1218                 /* Block error histogram and types */
1219                 int len;
1220                 unsigned int *percentiles = NULL;
1221                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1222
1223                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1224                                              ts->percentile_list,
1225                                              &percentiles, block_state_counts);
1226
1227                 if (len) {
1228                         struct json_object *block, *percentile_object, *states;
1229                         int state;
1230                         block = json_create_object();
1231                         json_object_add_value_object(root, "block", block);
1232
1233                         percentile_object = json_create_object();
1234                         json_object_add_value_object(block, "percentiles",
1235                                                      percentile_object);
1236                         for (i = 0; i < len; i++) {
1237                                 char buf[20];
1238                                 snprintf(buf, sizeof(buf), "%f",
1239                                          ts->percentile_list[i].u.f);
1240                                 json_object_add_value_int(percentile_object,
1241                                                           (const char *)buf,
1242                                                           percentiles[i]);
1243                         }
1244
1245                         states = json_create_object();
1246                         json_object_add_value_object(block, "states", states);
1247                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1248                                 json_object_add_value_int(states,
1249                                         block_state_names[state],
1250                                         block_state_counts[state]);
1251                         }
1252                         free(percentiles);
1253                 }
1254         }
1255
1256         return root;
1257 }
1258
1259 static void show_thread_status_terse(struct thread_stat *ts,
1260                                      struct group_run_stats *rs,
1261                                      struct buf_output *out)
1262 {
1263         if (terse_version == 2)
1264                 show_thread_status_terse_v2(ts, rs, out);
1265         else if (terse_version == 3 || terse_version == 4)
1266                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1267         else
1268                 log_err("fio: bad terse version!? %d\n", terse_version);
1269 }
1270
1271 struct json_object *show_thread_status(struct thread_stat *ts,
1272                                        struct group_run_stats *rs,
1273                                        struct flist_head *opt_list,
1274                                        struct buf_output *out)
1275 {
1276         struct json_object *ret = NULL;
1277
1278         if (output_format & FIO_OUTPUT_TERSE)
1279                 show_thread_status_terse(ts, rs,  out);
1280         if (output_format & FIO_OUTPUT_JSON)
1281                 ret = show_thread_status_json(ts, rs, opt_list);
1282         if (output_format & FIO_OUTPUT_NORMAL)
1283                 show_thread_status_normal(ts, rs,  out);
1284
1285         return ret;
1286 }
1287
1288 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1289 {
1290         double mean, S;
1291
1292         if (src->samples == 0)
1293                 return;
1294
1295         dst->min_val = min(dst->min_val, src->min_val);
1296         dst->max_val = max(dst->max_val, src->max_val);
1297
1298         /*
1299          * Compute new mean and S after the merge
1300          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1301          *  #Parallel_algorithm>
1302          */
1303         if (first) {
1304                 mean = src->mean.u.f;
1305                 S = src->S.u.f;
1306         } else {
1307                 double delta = src->mean.u.f - dst->mean.u.f;
1308
1309                 mean = ((src->mean.u.f * src->samples) +
1310                         (dst->mean.u.f * dst->samples)) /
1311                         (dst->samples + src->samples);
1312
1313                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1314                         (dst->samples * src->samples) /
1315                         (dst->samples + src->samples);
1316         }
1317
1318         dst->samples += src->samples;
1319         dst->mean.u.f = mean;
1320         dst->S.u.f = S;
1321 }
1322
1323 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1324 {
1325         int i;
1326
1327         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1328                 if (dst->max_run[i] < src->max_run[i])
1329                         dst->max_run[i] = src->max_run[i];
1330                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1331                         dst->min_run[i] = src->min_run[i];
1332                 if (dst->max_bw[i] < src->max_bw[i])
1333                         dst->max_bw[i] = src->max_bw[i];
1334                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1335                         dst->min_bw[i] = src->min_bw[i];
1336
1337                 dst->io_kb[i] += src->io_kb[i];
1338                 dst->agg[i] += src->agg[i];
1339         }
1340
1341         if (!dst->kb_base)
1342                 dst->kb_base = src->kb_base;
1343         if (!dst->unit_base)
1344                 dst->unit_base = src->unit_base;
1345 }
1346
1347 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1348                       bool first)
1349 {
1350         int l, k;
1351
1352         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1353                 if (!dst->unified_rw_rep) {
1354                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1355                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1356                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1357                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1358
1359                         dst->io_bytes[l] += src->io_bytes[l];
1360
1361                         if (dst->runtime[l] < src->runtime[l])
1362                                 dst->runtime[l] = src->runtime[l];
1363                 } else {
1364                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1365                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1366                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1367                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1368
1369                         dst->io_bytes[0] += src->io_bytes[l];
1370
1371                         if (dst->runtime[0] < src->runtime[l])
1372                                 dst->runtime[0] = src->runtime[l];
1373
1374                         /*
1375                          * We're summing to the same destination, so override
1376                          * 'first' after the first iteration of the loop
1377                          */
1378                         first = false;
1379                 }
1380         }
1381
1382         dst->usr_time += src->usr_time;
1383         dst->sys_time += src->sys_time;
1384         dst->ctx += src->ctx;
1385         dst->majf += src->majf;
1386         dst->minf += src->minf;
1387
1388         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1389                 dst->io_u_map[k] += src->io_u_map[k];
1390         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1391                 dst->io_u_submit[k] += src->io_u_submit[k];
1392         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1393                 dst->io_u_complete[k] += src->io_u_complete[k];
1394         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1395                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1396         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1397                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1398
1399         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1400                 if (!dst->unified_rw_rep) {
1401                         dst->total_io_u[k] += src->total_io_u[k];
1402                         dst->short_io_u[k] += src->short_io_u[k];
1403                         dst->drop_io_u[k] += src->drop_io_u[k];
1404                 } else {
1405                         dst->total_io_u[0] += src->total_io_u[k];
1406                         dst->short_io_u[0] += src->short_io_u[k];
1407                         dst->drop_io_u[0] += src->drop_io_u[k];
1408                 }
1409         }
1410
1411         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1412                 int m;
1413
1414                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1415                         if (!dst->unified_rw_rep)
1416                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1417                         else
1418                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1419                 }
1420         }
1421
1422         dst->total_run_time += src->total_run_time;
1423         dst->total_submit += src->total_submit;
1424         dst->total_complete += src->total_complete;
1425 }
1426
1427 void init_group_run_stat(struct group_run_stats *gs)
1428 {
1429         int i;
1430         memset(gs, 0, sizeof(*gs));
1431
1432         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1433                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1434 }
1435
1436 void init_thread_stat(struct thread_stat *ts)
1437 {
1438         int j;
1439
1440         memset(ts, 0, sizeof(*ts));
1441
1442         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1443                 ts->lat_stat[j].min_val = -1UL;
1444                 ts->clat_stat[j].min_val = -1UL;
1445                 ts->slat_stat[j].min_val = -1UL;
1446                 ts->bw_stat[j].min_val = -1UL;
1447         }
1448         ts->groupid = -1;
1449 }
1450
1451 void __show_run_stats(void)
1452 {
1453         struct group_run_stats *runstats, *rs;
1454         struct thread_data *td;
1455         struct thread_stat *threadstats, *ts;
1456         int i, j, k, nr_ts, last_ts, idx;
1457         int kb_base_warned = 0;
1458         int unit_base_warned = 0;
1459         struct json_object *root = NULL;
1460         struct json_array *array = NULL;
1461         struct buf_output output[FIO_OUTPUT_NR];
1462         struct flist_head **opt_lists;
1463
1464         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1465
1466         for (i = 0; i < groupid + 1; i++)
1467                 init_group_run_stat(&runstats[i]);
1468
1469         /*
1470          * find out how many threads stats we need. if group reporting isn't
1471          * enabled, it's one-per-td.
1472          */
1473         nr_ts = 0;
1474         last_ts = -1;
1475         for_each_td(td, i) {
1476                 if (!td->o.group_reporting) {
1477                         nr_ts++;
1478                         continue;
1479                 }
1480                 if (last_ts == td->groupid)
1481                         continue;
1482
1483                 last_ts = td->groupid;
1484                 nr_ts++;
1485         }
1486
1487         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1488         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1489
1490         for (i = 0; i < nr_ts; i++) {
1491                 init_thread_stat(&threadstats[i]);
1492                 opt_lists[i] = NULL;
1493         }
1494
1495         j = 0;
1496         last_ts = -1;
1497         idx = 0;
1498         for_each_td(td, i) {
1499                 if (idx && (!td->o.group_reporting ||
1500                     (td->o.group_reporting && last_ts != td->groupid))) {
1501                         idx = 0;
1502                         j++;
1503                 }
1504
1505                 last_ts = td->groupid;
1506
1507                 ts = &threadstats[j];
1508
1509                 ts->clat_percentiles = td->o.clat_percentiles;
1510                 ts->percentile_precision = td->o.percentile_precision;
1511                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1512                 opt_lists[j] = &td->opt_list;
1513
1514                 idx++;
1515                 ts->members++;
1516
1517                 if (ts->groupid == -1) {
1518                         /*
1519                          * These are per-group shared already
1520                          */
1521                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1522                         if (td->o.description)
1523                                 strncpy(ts->description, td->o.description,
1524                                                 FIO_JOBDESC_SIZE - 1);
1525                         else
1526                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1527
1528                         /*
1529                          * If multiple entries in this group, this is
1530                          * the first member.
1531                          */
1532                         ts->thread_number = td->thread_number;
1533                         ts->groupid = td->groupid;
1534
1535                         /*
1536                          * first pid in group, not very useful...
1537                          */
1538                         ts->pid = td->pid;
1539
1540                         ts->kb_base = td->o.kb_base;
1541                         ts->unit_base = td->o.unit_base;
1542                         ts->unified_rw_rep = td->o.unified_rw_rep;
1543                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1544                         log_info("fio: kb_base differs for jobs in group, using"
1545                                  " %u as the base\n", ts->kb_base);
1546                         kb_base_warned = 1;
1547                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1548                         log_info("fio: unit_base differs for jobs in group, using"
1549                                  " %u as the base\n", ts->unit_base);
1550                         unit_base_warned = 1;
1551                 }
1552
1553                 ts->continue_on_error = td->o.continue_on_error;
1554                 ts->total_err_count += td->total_err_count;
1555                 ts->first_error = td->first_error;
1556                 if (!ts->error) {
1557                         if (!td->error && td->o.continue_on_error &&
1558                             td->first_error) {
1559                                 ts->error = td->first_error;
1560                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1561                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1562                         } else  if (td->error) {
1563                                 ts->error = td->error;
1564                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1565                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1566                         }
1567                 }
1568
1569                 ts->latency_depth = td->latency_qd;
1570                 ts->latency_target = td->o.latency_target;
1571                 ts->latency_percentile = td->o.latency_percentile;
1572                 ts->latency_window = td->o.latency_window;
1573
1574                 ts->nr_block_infos = td->ts.nr_block_infos;
1575                 for (k = 0; k < ts->nr_block_infos; k++)
1576                         ts->block_infos[k] = td->ts.block_infos[k];
1577
1578                 sum_thread_stats(ts, &td->ts, idx == 1);
1579         }
1580
1581         for (i = 0; i < nr_ts; i++) {
1582                 unsigned long long bw;
1583
1584                 ts = &threadstats[i];
1585                 if (ts->groupid == -1)
1586                         continue;
1587                 rs = &runstats[ts->groupid];
1588                 rs->kb_base = ts->kb_base;
1589                 rs->unit_base = ts->unit_base;
1590                 rs->unified_rw_rep += ts->unified_rw_rep;
1591
1592                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1593                         if (!ts->runtime[j])
1594                                 continue;
1595                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1596                                 rs->min_run[j] = ts->runtime[j];
1597                         if (ts->runtime[j] > rs->max_run[j])
1598                                 rs->max_run[j] = ts->runtime[j];
1599
1600                         bw = 0;
1601                         if (ts->runtime[j]) {
1602                                 unsigned long runt = ts->runtime[j];
1603                                 unsigned long long kb;
1604
1605                                 kb = ts->io_bytes[j] / rs->kb_base;
1606                                 bw = kb * 1000 / runt;
1607                         }
1608                         if (bw < rs->min_bw[j])
1609                                 rs->min_bw[j] = bw;
1610                         if (bw > rs->max_bw[j])
1611                                 rs->max_bw[j] = bw;
1612
1613                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1614                 }
1615         }
1616
1617         for (i = 0; i < groupid + 1; i++) {
1618                 int ddir;
1619
1620                 rs = &runstats[i];
1621
1622                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1623                         if (rs->max_run[ddir])
1624                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1625                                                 rs->max_run[ddir];
1626                 }
1627         }
1628
1629         for (i = 0; i < FIO_OUTPUT_NR; i++)
1630                 buf_output_init(&output[i]);
1631
1632         /*
1633          * don't overwrite last signal output
1634          */
1635         if (output_format & FIO_OUTPUT_NORMAL)
1636                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1637         if (output_format & FIO_OUTPUT_JSON) {
1638                 struct thread_data *global;
1639                 char time_buf[32];
1640                 struct timeval now;
1641                 unsigned long long ms_since_epoch;
1642
1643                 gettimeofday(&now, NULL);
1644                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1645                                  (unsigned long long)(now.tv_usec) / 1000;
1646
1647                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1648                                 sizeof(time_buf));
1649                 time_buf[strlen(time_buf) - 1] = '\0';
1650
1651                 root = json_create_object();
1652                 json_object_add_value_string(root, "fio version", fio_version_string);
1653                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1654                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1655                 json_object_add_value_string(root, "time", time_buf);
1656                 global = get_global_options();
1657                 json_add_job_opts(root, "global options", &global->opt_list, false);
1658                 array = json_create_array();
1659                 json_object_add_value_array(root, "jobs", array);
1660         }
1661
1662         if (is_backend)
1663                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1664
1665         for (i = 0; i < nr_ts; i++) {
1666                 ts = &threadstats[i];
1667                 rs = &runstats[ts->groupid];
1668
1669                 if (is_backend) {
1670                         fio_server_send_job_options(opt_lists[i], i);
1671                         fio_server_send_ts(ts, rs);
1672                 } else {
1673                         if (output_format & FIO_OUTPUT_TERSE)
1674                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1675                         if (output_format & FIO_OUTPUT_JSON) {
1676                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1677                                 json_array_add_value_object(array, tmp);
1678                         }
1679                         if (output_format & FIO_OUTPUT_NORMAL)
1680                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1681                 }
1682         }
1683         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1684                 /* disk util stats, if any */
1685                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1686
1687                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1688
1689                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1690                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1691                 json_free_object(root);
1692         }
1693
1694         for (i = 0; i < groupid + 1; i++) {
1695                 rs = &runstats[i];
1696
1697                 rs->groupid = i;
1698                 if (is_backend)
1699                         fio_server_send_gs(rs);
1700                 else if (output_format & FIO_OUTPUT_NORMAL)
1701                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1702         }
1703
1704         if (is_backend)
1705                 fio_server_send_du();
1706         else if (output_format & FIO_OUTPUT_NORMAL) {
1707                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1708                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1709         }
1710
1711         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1712                 buf_output_flush(&output[i]);
1713                 buf_output_free(&output[i]);
1714         }
1715
1716         log_info_flush();
1717         free(runstats);
1718         free(threadstats);
1719         free(opt_lists);
1720 }
1721
1722 void show_run_stats(void)
1723 {
1724         fio_mutex_down(stat_mutex);
1725         __show_run_stats();
1726         fio_mutex_up(stat_mutex);
1727 }
1728
1729 void __show_running_run_stats(void)
1730 {
1731         struct thread_data *td;
1732         unsigned long long *rt;
1733         struct timeval tv;
1734         int i;
1735
1736         fio_mutex_down(stat_mutex);
1737
1738         rt = malloc(thread_number * sizeof(unsigned long long));
1739         fio_gettime(&tv, NULL);
1740
1741         for_each_td(td, i) {
1742                 td->update_rusage = 1;
1743                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1744                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1745                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1746                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1747
1748                 rt[i] = mtime_since(&td->start, &tv);
1749                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1750                         td->ts.runtime[DDIR_READ] += rt[i];
1751                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1752                         td->ts.runtime[DDIR_WRITE] += rt[i];
1753                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1754                         td->ts.runtime[DDIR_TRIM] += rt[i];
1755         }
1756
1757         for_each_td(td, i) {
1758                 if (td->runstate >= TD_EXITED)
1759                         continue;
1760                 if (td->rusage_sem) {
1761                         td->update_rusage = 1;
1762                         fio_mutex_down(td->rusage_sem);
1763                 }
1764                 td->update_rusage = 0;
1765         }
1766
1767         __show_run_stats();
1768
1769         for_each_td(td, i) {
1770                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1771                         td->ts.runtime[DDIR_READ] -= rt[i];
1772                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1773                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1774                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1775                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1776         }
1777
1778         free(rt);
1779         fio_mutex_up(stat_mutex);
1780 }
1781
1782 static int status_interval_init;
1783 static struct timeval status_time;
1784 static int status_file_disabled;
1785
1786 #define FIO_STATUS_FILE         "fio-dump-status"
1787
1788 static int check_status_file(void)
1789 {
1790         struct stat sb;
1791         const char *temp_dir;
1792         char fio_status_file_path[PATH_MAX];
1793
1794         if (status_file_disabled)
1795                 return 0;
1796
1797         temp_dir = getenv("TMPDIR");
1798         if (temp_dir == NULL) {
1799                 temp_dir = getenv("TEMP");
1800                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1801                         temp_dir = NULL;
1802         }
1803         if (temp_dir == NULL)
1804                 temp_dir = "/tmp";
1805
1806         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1807
1808         if (stat(fio_status_file_path, &sb))
1809                 return 0;
1810
1811         if (unlink(fio_status_file_path) < 0) {
1812                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1813                                                         strerror(errno));
1814                 log_err("fio: disabling status file updates\n");
1815                 status_file_disabled = 1;
1816         }
1817
1818         return 1;
1819 }
1820
1821 void check_for_running_stats(void)
1822 {
1823         if (status_interval) {
1824                 if (!status_interval_init) {
1825                         fio_gettime(&status_time, NULL);
1826                         status_interval_init = 1;
1827                 } else if (mtime_since_now(&status_time) >= status_interval) {
1828                         show_running_run_stats();
1829                         fio_gettime(&status_time, NULL);
1830                         return;
1831                 }
1832         }
1833         if (check_status_file()) {
1834                 show_running_run_stats();
1835                 return;
1836         }
1837 }
1838
1839 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1840 {
1841         double val = data;
1842         double delta;
1843
1844         if (data > is->max_val)
1845                 is->max_val = data;
1846         if (data < is->min_val)
1847                 is->min_val = data;
1848
1849         delta = val - is->mean.u.f;
1850         if (delta) {
1851                 is->mean.u.f += delta / (is->samples + 1.0);
1852                 is->S.u.f += delta * (val - is->mean.u.f);
1853         }
1854
1855         is->samples++;
1856 }
1857
1858 /*
1859  * Return a struct io_logs, which is added to the tail of the log
1860  * list for 'iolog'.
1861  */
1862 static struct io_logs *get_new_log(struct io_log *iolog)
1863 {
1864         size_t new_size, new_samples;
1865         struct io_logs *cur_log;
1866
1867         /*
1868          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1869          * forever
1870          */
1871         if (!iolog->cur_log_max)
1872                 new_samples = DEF_LOG_ENTRIES;
1873         else {
1874                 new_samples = iolog->cur_log_max * 2;
1875                 if (new_samples > MAX_LOG_ENTRIES)
1876                         new_samples = MAX_LOG_ENTRIES;
1877         }
1878
1879         new_size = new_samples * log_entry_sz(iolog);
1880
1881         cur_log = smalloc(sizeof(*cur_log));
1882         if (cur_log) {
1883                 INIT_FLIST_HEAD(&cur_log->list);
1884                 cur_log->log = malloc(new_size);
1885                 if (cur_log->log) {
1886                         cur_log->nr_samples = 0;
1887                         cur_log->max_samples = new_samples;
1888                         flist_add_tail(&cur_log->list, &iolog->io_logs);
1889                         iolog->cur_log_max = new_samples;
1890                         return cur_log;
1891                 }
1892                 sfree(cur_log);
1893         }
1894
1895         return NULL;
1896 }
1897
1898 /*
1899  * Add and return a new log chunk, or return current log if big enough
1900  */
1901 static struct io_logs *regrow_log(struct io_log *iolog)
1902 {
1903         struct io_logs *cur_log;
1904         int i;
1905
1906         if (!iolog || iolog->disabled)
1907                 goto disable;
1908
1909         cur_log = iolog_cur_log(iolog);
1910         if (!cur_log) {
1911                 cur_log = get_new_log(iolog);
1912                 if (!cur_log)
1913                         return NULL;
1914         }
1915
1916         if (cur_log->nr_samples < cur_log->max_samples)
1917                 return cur_log;
1918
1919         /*
1920          * No room for a new sample. If we're compressing on the fly, flush
1921          * out the current chunk
1922          */
1923         if (iolog->log_gz) {
1924                 if (iolog_cur_flush(iolog, cur_log)) {
1925                         log_err("fio: failed flushing iolog! Will stop logging.\n");
1926                         return NULL;
1927                 }
1928         }
1929
1930         /*
1931          * Get a new log array, and add to our list
1932          */
1933         cur_log = get_new_log(iolog);
1934         if (!cur_log) {
1935                 log_err("fio: failed extending iolog! Will stop logging.\n");
1936                 return NULL;
1937         }
1938
1939         if (!iolog->pending || !iolog->pending->nr_samples)
1940                 return cur_log;
1941
1942         /*
1943          * Flush pending items to new log
1944          */
1945         for (i = 0; i < iolog->pending->nr_samples; i++) {
1946                 struct io_sample *src, *dst;
1947
1948                 src = get_sample(iolog, iolog->pending, i);
1949                 dst = get_sample(iolog, cur_log, i);
1950                 memcpy(dst, src, log_entry_sz(iolog));
1951         }
1952         cur_log->nr_samples = iolog->pending->nr_samples;
1953
1954         iolog->pending->nr_samples = 0;
1955         return cur_log;
1956 disable:
1957         if (iolog)
1958                 iolog->disabled = true;
1959         return NULL;
1960 }
1961
1962 void regrow_logs(struct thread_data *td)
1963 {
1964         regrow_log(td->slat_log);
1965         regrow_log(td->clat_log);
1966         regrow_log(td->lat_log);
1967         regrow_log(td->bw_log);
1968         regrow_log(td->iops_log);
1969         td->flags &= ~TD_F_REGROW_LOGS;
1970 }
1971
1972 static struct io_logs *get_cur_log(struct io_log *iolog)
1973 {
1974         struct io_logs *cur_log;
1975
1976         cur_log = iolog_cur_log(iolog);
1977         if (!cur_log) {
1978                 cur_log = get_new_log(iolog);
1979                 if (!cur_log)
1980                         return NULL;
1981         }
1982
1983         if (cur_log->nr_samples < cur_log->max_samples)
1984                 return cur_log;
1985
1986         /*
1987          * Out of space. If we're in IO offload mode, or we're not doing
1988          * per unit logging (hence logging happens outside of the IO thread
1989          * as well), add a new log chunk inline. If we're doing inline
1990          * submissions, flag 'td' as needing a log regrow and we'll take
1991          * care of it on the submission side.
1992          */
1993         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
1994             !per_unit_log(iolog))
1995                 return regrow_log(iolog);
1996
1997         iolog->td->flags |= TD_F_REGROW_LOGS;
1998         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
1999         return iolog->pending;
2000 }
2001
2002 static void __add_log_sample(struct io_log *iolog, unsigned long val,
2003                              enum fio_ddir ddir, unsigned int bs,
2004                              unsigned long t, uint64_t offset)
2005 {
2006         struct io_logs *cur_log;
2007
2008         if (iolog->disabled)
2009                 return;
2010         if (flist_empty(&iolog->io_logs))
2011                 iolog->avg_last = t;
2012
2013         cur_log = get_cur_log(iolog);
2014         if (cur_log) {
2015                 struct io_sample *s;
2016
2017                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2018
2019                 s->val = val;
2020                 s->time = t;
2021                 io_sample_set_ddir(iolog, s, ddir);
2022                 s->bs = bs;
2023
2024                 if (iolog->log_offset) {
2025                         struct io_sample_offset *so = (void *) s;
2026
2027                         so->offset = offset;
2028                 }
2029
2030                 cur_log->nr_samples++;
2031                 return;
2032         }
2033
2034         iolog->disabled = true;
2035 }
2036
2037 static inline void reset_io_stat(struct io_stat *ios)
2038 {
2039         ios->max_val = ios->min_val = ios->samples = 0;
2040         ios->mean.u.f = ios->S.u.f = 0;
2041 }
2042
2043 void reset_io_stats(struct thread_data *td)
2044 {
2045         struct thread_stat *ts = &td->ts;
2046         int i, j;
2047
2048         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2049                 reset_io_stat(&ts->clat_stat[i]);
2050                 reset_io_stat(&ts->slat_stat[i]);
2051                 reset_io_stat(&ts->lat_stat[i]);
2052                 reset_io_stat(&ts->bw_stat[i]);
2053                 reset_io_stat(&ts->iops_stat[i]);
2054
2055                 ts->io_bytes[i] = 0;
2056                 ts->runtime[i] = 0;
2057
2058                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2059                         ts->io_u_plat[i][j] = 0;
2060         }
2061
2062         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2063                 ts->io_u_map[i] = 0;
2064                 ts->io_u_submit[i] = 0;
2065                 ts->io_u_complete[i] = 0;
2066                 ts->io_u_lat_u[i] = 0;
2067                 ts->io_u_lat_m[i] = 0;
2068                 ts->total_submit = 0;
2069                 ts->total_complete = 0;
2070         }
2071
2072         for (i = 0; i < 3; i++) {
2073                 ts->total_io_u[i] = 0;
2074                 ts->short_io_u[i] = 0;
2075                 ts->drop_io_u[i] = 0;
2076         }
2077 }
2078
2079 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2080                               unsigned long elapsed, bool log_max)
2081 {
2082         /*
2083          * Note an entry in the log. Use the mean from the logged samples,
2084          * making sure to properly round up. Only write a log entry if we
2085          * had actual samples done.
2086          */
2087         if (iolog->avg_window[ddir].samples) {
2088                 unsigned long val;
2089
2090                 if (log_max)
2091                         val = iolog->avg_window[ddir].max_val;
2092                 else
2093                         val = iolog->avg_window[ddir].mean.u.f + 0.50;
2094
2095                 __add_log_sample(iolog, val, ddir, 0, elapsed, 0);
2096         }
2097
2098         reset_io_stat(&iolog->avg_window[ddir]);
2099 }
2100
2101 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2102                              bool log_max)
2103 {
2104         int ddir;
2105
2106         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2107                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2108 }
2109
2110 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
2111                            unsigned long val, enum fio_ddir ddir,
2112                            unsigned int bs, uint64_t offset)
2113 {
2114         unsigned long elapsed, this_window;
2115
2116         if (!ddir_rw(ddir))
2117                 return;
2118
2119         elapsed = mtime_since_now(&td->epoch);
2120
2121         /*
2122          * If no time averaging, just add the log sample.
2123          */
2124         if (!iolog->avg_msec) {
2125                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
2126                 return;
2127         }
2128
2129         /*
2130          * Add the sample. If the time period has passed, then
2131          * add that entry to the log and clear.
2132          */
2133         add_stat_sample(&iolog->avg_window[ddir], val);
2134
2135         /*
2136          * If period hasn't passed, adding the above sample is all we
2137          * need to do.
2138          */
2139         this_window = elapsed - iolog->avg_last;
2140         if (this_window < iolog->avg_msec)
2141                 return;
2142
2143         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2144
2145         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2146 }
2147
2148 void finalize_logs(struct thread_data *td, bool unit_logs)
2149 {
2150         unsigned long elapsed;
2151
2152         elapsed = mtime_since_now(&td->epoch);
2153
2154         if (td->clat_log && unit_logs)
2155                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2156         if (td->slat_log && unit_logs)
2157                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2158         if (td->lat_log && unit_logs)
2159                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2160         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2161                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2162         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2163                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2164 }
2165
2166 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
2167 {
2168         struct io_log *iolog;
2169
2170         if (!ddir_rw(ddir))
2171                 return;
2172
2173         iolog = agg_io_log[ddir];
2174         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
2175 }
2176
2177 static void add_clat_percentile_sample(struct thread_stat *ts,
2178                                 unsigned long usec, enum fio_ddir ddir)
2179 {
2180         unsigned int idx = plat_val_to_idx(usec);
2181         assert(idx < FIO_IO_U_PLAT_NR);
2182
2183         ts->io_u_plat[ddir][idx]++;
2184 }
2185
2186 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2187                      unsigned long usec, unsigned int bs, uint64_t offset)
2188 {
2189         struct thread_stat *ts = &td->ts;
2190
2191         td_io_u_lock(td);
2192
2193         add_stat_sample(&ts->clat_stat[ddir], usec);
2194
2195         if (td->clat_log)
2196                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
2197
2198         if (ts->clat_percentiles)
2199                 add_clat_percentile_sample(ts, usec, ddir);
2200
2201         td_io_u_unlock(td);
2202 }
2203
2204 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2205                      unsigned long usec, unsigned int bs, uint64_t offset)
2206 {
2207         struct thread_stat *ts = &td->ts;
2208
2209         if (!ddir_rw(ddir))
2210                 return;
2211
2212         td_io_u_lock(td);
2213
2214         add_stat_sample(&ts->slat_stat[ddir], usec);
2215
2216         if (td->slat_log)
2217                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2218
2219         td_io_u_unlock(td);
2220 }
2221
2222 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2223                     unsigned long usec, unsigned int bs, uint64_t offset)
2224 {
2225         struct thread_stat *ts = &td->ts;
2226
2227         if (!ddir_rw(ddir))
2228                 return;
2229
2230         td_io_u_lock(td);
2231
2232         add_stat_sample(&ts->lat_stat[ddir], usec);
2233
2234         if (td->lat_log)
2235                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2236
2237         td_io_u_unlock(td);
2238 }
2239
2240 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2241                    unsigned int bytes, unsigned long spent)
2242 {
2243         struct thread_stat *ts = &td->ts;
2244         unsigned long rate;
2245
2246         if (spent)
2247                 rate = bytes * 1000 / spent;
2248         else
2249                 rate = 0;
2250
2251         td_io_u_lock(td);
2252
2253         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2254
2255         if (td->bw_log)
2256                 add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
2257
2258         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2259         td_io_u_unlock(td);
2260 }
2261
2262 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2263 {
2264         struct thread_stat *ts = &td->ts;
2265         unsigned long spent, rate;
2266         enum fio_ddir ddir;
2267
2268         spent = mtime_since(&td->bw_sample_time, t);
2269         if (spent < td->o.bw_avg_time &&
2270             td->o.bw_avg_time - spent >= 10)
2271                 return td->o.bw_avg_time - spent;
2272
2273         td_io_u_lock(td);
2274
2275         /*
2276          * Compute both read and write rates for the interval.
2277          */
2278         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2279                 uint64_t delta;
2280
2281                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2282                 if (!delta)
2283                         continue; /* No entries for interval */
2284
2285                 if (spent)
2286                         rate = delta * 1000 / spent / 1024;
2287                 else
2288                         rate = 0;
2289
2290                 add_stat_sample(&ts->bw_stat[ddir], rate);
2291
2292                 if (td->bw_log) {
2293                         unsigned int bs = 0;
2294
2295                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2296                                 bs = td->o.min_bs[ddir];
2297
2298                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2299                 }
2300
2301                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2302         }
2303
2304         timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
2305
2306         td_io_u_unlock(td);
2307
2308         if (spent <= td->o.bw_avg_time)
2309                 return td->o.bw_avg_time;
2310
2311         return td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
2312 }
2313
2314 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2315                      unsigned int bytes)
2316 {
2317         struct thread_stat *ts = &td->ts;
2318
2319         td_io_u_lock(td);
2320
2321         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2322
2323         if (td->iops_log)
2324                 add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
2325
2326         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2327         td_io_u_unlock(td);
2328 }
2329
2330 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2331 {
2332         struct thread_stat *ts = &td->ts;
2333         unsigned long spent, iops;
2334         enum fio_ddir ddir;
2335
2336         spent = mtime_since(&td->iops_sample_time, t);
2337         if (spent < td->o.iops_avg_time &&
2338             td->o.iops_avg_time - spent >= 10)
2339                 return td->o.iops_avg_time - spent;
2340
2341         td_io_u_lock(td);
2342
2343         /*
2344          * Compute both read and write rates for the interval.
2345          */
2346         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2347                 uint64_t delta;
2348
2349                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2350                 if (!delta)
2351                         continue; /* No entries for interval */
2352
2353                 if (spent)
2354                         iops = (delta * 1000) / spent;
2355                 else
2356                         iops = 0;
2357
2358                 add_stat_sample(&ts->iops_stat[ddir], iops);
2359
2360                 if (td->iops_log) {
2361                         unsigned int bs = 0;
2362
2363                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2364                                 bs = td->o.min_bs[ddir];
2365
2366                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2367                 }
2368
2369                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2370         }
2371
2372         timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
2373
2374         td_io_u_unlock(td);
2375
2376         if (spent <= td->o.iops_avg_time)
2377                 return td->o.iops_avg_time;
2378
2379         return td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
2380 }
2381
2382 /*
2383  * Returns msecs to next event
2384  */
2385 int calc_log_samples(void)
2386 {
2387         struct thread_data *td;
2388         unsigned int next = ~0U, tmp;
2389         struct timeval now;
2390         int i;
2391
2392         fio_gettime(&now, NULL);
2393
2394         for_each_td(td, i) {
2395                 if (in_ramp_time(td) ||
2396                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2397                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2398                         continue;
2399                 }
2400                 if (!per_unit_log(td->bw_log)) {
2401                         tmp = add_bw_samples(td, &now);
2402                         if (tmp < next)
2403                                 next = tmp;
2404                 }
2405                 if (!per_unit_log(td->iops_log)) {
2406                         tmp = add_iops_samples(td, &now);
2407                         if (tmp < next)
2408                                 next = tmp;
2409                 }
2410         }
2411
2412         return next == ~0U ? 0 : next;
2413 }
2414
2415 void stat_init(void)
2416 {
2417         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2418 }
2419
2420 void stat_exit(void)
2421 {
2422         /*
2423          * When we have the mutex, we know out-of-band access to it
2424          * have ended.
2425          */
2426         fio_mutex_down(stat_mutex);
2427         fio_mutex_remove(stat_mutex);
2428 }
2429
2430 /*
2431  * Called from signal handler. Wake up status thread.
2432  */
2433 void show_running_run_stats(void)
2434 {
2435         helper_do_stat();
2436 }
2437
2438 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2439 {
2440         /* Ignore io_u's which span multiple blocks--they will just get
2441          * inaccurate counts. */
2442         int idx = (io_u->offset - io_u->file->file_offset)
2443                         / td->o.bs[DDIR_TRIM];
2444         uint32_t *info = &td->ts.block_infos[idx];
2445         assert(idx < td->ts.nr_block_infos);
2446         return info;
2447 }