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