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