stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18
  19 struct fio_mutex *stat_mutex;
  20
  21 void update_rusage_stat(struct thread_data *td)
  22 {
  23         struct thread_stat *ts = &td->ts;
  24
  25         fio_getrusage(&td->ru_end);
  26         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  27                                         &td->ru_end.ru_utime);
  28         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  29                                         &td->ru_end.ru_stime);
  30         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  31                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  32         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  33         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  34
  35         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  36 }
  37
  38 /*
  39  * Given a latency, return the index of the corresponding bucket in
  40  * the structure tracking percentiles.
  41  *
  42  * (1) find the group (and error bits) that the value (latency)
  43  * belongs to by looking at its MSB. (2) find the bucket number in the
  44  * group by looking at the index bits.
  45  *
  46  */
  47 static unsigned int plat_val_to_idx(unsigned int val)
  48 {
  49         unsigned int msb, error_bits, base, offset, idx;
  50
  51         /* Find MSB starting from bit 0 */
  52         if (val == 0)
  53                 msb = 0;
  54         else
  55                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  56
  57         /*
  58          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  59          * all bits of the sample as index
  60          */
  61         if (msb <= FIO_IO_U_PLAT_BITS)
  62                 return val;
  63
  64         /* Compute the number of error bits to discard*/
  65         error_bits = msb - FIO_IO_U_PLAT_BITS;
  66
  67         /* Compute the number of buckets before the group */
  68         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  69
  70         /*
  71          * Discard the error bits and apply the mask to find the
  72          * index for the buckets in the group
  73          */
  74         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  75
  76         /* Make sure the index does not exceed (array size - 1) */
  77         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  78                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  79
  80         return idx;
  81 }
  82
  83 /*
  84  * Convert the given index of the bucket array to the value
  85  * represented by the bucket
  86  */
  87 static unsigned int plat_idx_to_val(unsigned int idx)
  88 {
  89         unsigned int error_bits, k, base;
  90
  91         assert(idx < FIO_IO_U_PLAT_NR);
  92
  93         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  94          * all bits of the sample as index */
  95         if (idx < (FIO_IO_U_PLAT_VAL << 1))
  96                 return idx;
  97
  98         /* Find the group and compute the minimum value of that group */
  99         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 100         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 101
 102         /* Find its bucket number of the group */
 103         k = idx % FIO_IO_U_PLAT_VAL;
 104
 105         /* Return the mean of the range of the bucket */
 106         return base + ((k + 0.5) * (1 << error_bits));
 107 }
 108
 109 static int double_cmp(const void *a, const void *b)
 110 {
 111         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 112         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 113         int cmp = 0;
 114
 115         if (fa.u.f > fb.u.f)
 116                 cmp = 1;
 117         else if (fa.u.f < fb.u.f)
 118                 cmp = -1;
 119
 120         return cmp;
 121 }
 122
 123 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 124                                    fio_fp64_t *plist, unsigned int **output,
 125                                    unsigned int *maxv, unsigned int *minv)
 126 {
 127         unsigned long sum = 0;
 128         unsigned int len, i, j = 0;
 129         unsigned int oval_len = 0;
 130         unsigned int *ovals = NULL;
 131         int is_last;
 132
 133         *minv = -1U;
 134         *maxv = 0;
 135
 136         len = 0;
 137         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 138                 len++;
 139
 140         if (!len)
 141                 return 0;
 142
 143         /*
 144          * Sort the percentile list. Note that it may already be sorted if
 145          * we are using the default values, but since it's a short list this
 146          * isn't a worry. Also note that this does not work for NaN values.
 147          */
 148         if (len > 1)
 149                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 150
 151         /*
 152          * Calculate bucket values, note down max and min values
 153          */
 154         is_last = 0;
 155         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 156                 sum += io_u_plat[i];
 157                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 158                         assert(plist[j].u.f <= 100.0);
 159
 160                         if (j == oval_len) {
 161                                 oval_len += 100;
 162                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 163                         }
 164
 165                         ovals[j] = plat_idx_to_val(i);
 166                         if (ovals[j] < *minv)
 167                                 *minv = ovals[j];
 168                         if (ovals[j] > *maxv)
 169                                 *maxv = ovals[j];
 170
 171                         is_last = (j == len - 1);
 172                         if (is_last)
 173                                 break;
 174
 175                         j++;
 176                 }
 177         }
 178
 179         *output = ovals;
 180         return len;
 181 }
 182
 183 /*
 184  * Find and display the p-th percentile of clat
 185  */
 186 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 187                                   fio_fp64_t *plist, unsigned int precision,
 188                                   struct buf_output *out)
 189 {
 190         unsigned int len, j = 0, minv, maxv;
 191         unsigned int *ovals;
 192         int is_last, per_line, scale_down;
 193         char fmt[32];
 194
 195         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 196         if (!len)
 197                 goto out;
 198
 199         /*
 200          * We default to usecs, but if the value range is such that we
 201          * should scale down to msecs, do that.
 202          */
 203         if (minv > 2000 && maxv > 99999) {
 204                 scale_down = 1;
 205                 log_buf(out, "    clat percentiles (msec):\n     |");
 206         } else {
 207                 scale_down = 0;
 208                 log_buf(out, "    clat percentiles (usec):\n     |");
 209         }
 210
 211         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 212         per_line = (80 - 7) / (precision + 14);
 213
 214         for (j = 0; j < len; j++) {
 215                 char fbuf[16], *ptr = fbuf;
 216
 217                 /* for formatting */
 218                 if (j != 0 && (j % per_line) == 0)
 219                         log_buf(out, "     |");
 220
 221                 /* end of the list */
 222                 is_last = (j == len - 1);
 223
 224                 if (plist[j].u.f < 10.0)
 225                         ptr += sprintf(fbuf, " ");
 226
 227                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 228
 229                 if (scale_down)
 230                         ovals[j] = (ovals[j] + 999) / 1000;
 231
 232                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 233
 234                 if (is_last)
 235                         break;
 236
 237                 if ((j % per_line) == per_line - 1)     /* for formatting */
 238                         log_buf(out, "\n");
 239         }
 240
 241 out:
 242         if (ovals)
 243                 free(ovals);
 244 }
 245
 246 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 247              double *mean, double *dev)
 248 {
 249         double n = (double) is->samples;
 250
 251         if (n == 0)
 252                 return 0;
 253
 254         *min = is->min_val;
 255         *max = is->max_val;
 256         *mean = is->mean.u.f;
 257
 258         if (n > 1.0)
 259                 *dev = sqrt(is->S.u.f / (n - 1.0));
 260         else
 261                 *dev = 0;
 262
 263         return 1;
 264 }
 265
 266 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 267 {
 268         char *p1, *p2, *p3, *p4;
 269         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 270         int i;
 271
 272         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 273
 274         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 275                 const int i2p = is_power_of_2(rs->kb_base);
 276
 277                 if (!rs->max_run[i])
 278                         continue;
 279
 280                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 281                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 282                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 283                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 284
 285                 log_buf(out, "%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 286                          " mint=%llumsec, maxt=%llumsec\n",
 287                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 288                                 p1, p2, p3, p4,
 289                                 (unsigned long long) rs->min_run[i],
 290                                 (unsigned long long) rs->max_run[i]);
 291
 292                 free(p1);
 293                 free(p2);
 294                 free(p3);
 295                 free(p4);
 296         }
 297 }
 298
 299 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 300 {
 301         int i;
 302
 303         /*
 304          * Do depth distribution calculations
 305          */
 306         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 307                 if (total) {
 308                         io_u_dist[i] = (double) map[i] / (double) total;
 309                         io_u_dist[i] *= 100.0;
 310                         if (io_u_dist[i] < 0.1 && map[i])
 311                                 io_u_dist[i] = 0.1;
 312                 } else
 313                         io_u_dist[i] = 0.0;
 314         }
 315 }
 316
 317 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 318                           unsigned int *src, int nr)
 319 {
 320         unsigned long total = ddir_rw_sum(ts->total_io_u);
 321         int i;
 322
 323         /*
 324          * Do latency distribution calculations
 325          */
 326         for (i = 0; i < nr; i++) {
 327                 if (total) {
 328                         dst[i] = (double) src[i] / (double) total;
 329                         dst[i] *= 100.0;
 330                         if (dst[i] < 0.01 && src[i])
 331                                 dst[i] = 0.01;
 332                 } else
 333                         dst[i] = 0.0;
 334         }
 335 }
 336
 337 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 338 {
 339         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 340 }
 341
 342 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 343 {
 344         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 345 }
 346
 347 static void display_lat(const char *name, unsigned long min, unsigned long max,
 348                         double mean, double dev, struct buf_output *out)
 349 {
 350         const char *base = "(usec)";
 351         char *minp, *maxp;
 352
 353         if (!usec_to_msec(&min, &max, &mean, &dev))
 354                 base = "(msec)";
 355
 356         minp = num2str(min, 6, 1, 0, 0);
 357         maxp = num2str(max, 6, 1, 0, 0);
 358
 359         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 360                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 361
 362         free(minp);
 363         free(maxp);
 364 }
 365
 366 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 367                              int ddir, struct buf_output *out)
 368 {
 369         const char *str[] = { "read ", "write", "trim" };
 370         unsigned long min, max, runt;
 371         unsigned long long bw, iops;
 372         double mean, dev;
 373         char *io_p, *bw_p, *iops_p;
 374         int i2p;
 375
 376         assert(ddir_rw(ddir));
 377
 378         if (!ts->runtime[ddir])
 379                 return;
 380
 381         i2p = is_power_of_2(rs->kb_base);
 382         runt = ts->runtime[ddir];
 383
 384         bw = (1000 * ts->io_bytes[ddir]) / runt;
 385         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 386         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 387
 388         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 389         iops_p = num2str(iops, 6, 1, 0, 0);
 390
 391         log_buf(out, "  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 392                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 393                                 io_p, bw_p, iops_p,
 394                                 (unsigned long long) ts->runtime[ddir]);
 395
 396         free(io_p);
 397         free(bw_p);
 398         free(iops_p);
 399
 400         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 401                 display_lat("slat", min, max, mean, dev, out);
 402         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 403                 display_lat("clat", min, max, mean, dev, out);
 404         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 405                 display_lat(" lat", min, max, mean, dev, out);
 406
 407         if (ts->clat_percentiles) {
 408                 show_clat_percentiles(ts->io_u_plat[ddir],
 409                                         ts->clat_stat[ddir].samples,
 410                                         ts->percentile_list,
 411                                         ts->percentile_precision, out);
 412         }
 413         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 414                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 415                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 416
 417                 if (rs->unit_base == 1) {
 418                         min *= 8.0;
 419                         max *= 8.0;
 420                         mean *= 8.0;
 421                         dev *= 8.0;
 422                 }
 423
 424                 if (rs->agg[ddir]) {
 425                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 426                         if (p_of_agg > 100.0)
 427                                 p_of_agg = 100.0;
 428                 }
 429
 430                 if (mean > fkb_base * fkb_base) {
 431                         min /= fkb_base;
 432                         max /= fkb_base;
 433                         mean /= fkb_base;
 434                         dev /= fkb_base;
 435                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 436                 }
 437
 438                 log_buf(out, "    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 439                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 440                                                         p_of_agg, mean, dev);
 441         }
 442 }
 443
 444 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 445                     const char *msg, struct buf_output *out)
 446 {
 447         int new_line = 1, i, line = 0, shown = 0;
 448
 449         for (i = 0; i < nr; i++) {
 450                 if (io_u_lat[i] <= 0.0)
 451                         continue;
 452                 shown = 1;
 453                 if (new_line) {
 454                         if (line)
 455                                 log_buf(out, "\n");
 456                         log_buf(out, "    lat (%s) : ", msg);
 457                         new_line = 0;
 458                         line = 0;
 459                 }
 460                 if (line)
 461                         log_buf(out, ", ");
 462                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 463                 line++;
 464                 if (line == 5)
 465                         new_line = 1;
 466         }
 467
 468         if (shown)
 469                 log_buf(out, "\n");
 470
 471         return shown;
 472 }
 473
 474 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 475 {
 476         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 477                                  "250=", "500=", "750=", "1000=", };
 478
 479         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 480 }
 481
 482 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 483 {
 484         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 485                                  "250=", "500=", "750=", "1000=", "2000=",
 486                                  ">=2000=", };
 487
 488         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 489 }
 490
 491 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 492 {
 493         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 494         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 495
 496         stat_calc_lat_u(ts, io_u_lat_u);
 497         stat_calc_lat_m(ts, io_u_lat_m);
 498
 499         show_lat_u(io_u_lat_u, out);
 500         show_lat_m(io_u_lat_m, out);
 501 }
 502
 503 static int block_state_category(int block_state)
 504 {
 505         switch (block_state) {
 506         case BLOCK_STATE_UNINIT:
 507                 return 0;
 508         case BLOCK_STATE_TRIMMED:
 509         case BLOCK_STATE_WRITTEN:
 510                 return 1;
 511         case BLOCK_STATE_WRITE_FAILURE:
 512         case BLOCK_STATE_TRIM_FAILURE:
 513                 return 2;
 514         default:
 515                 /* Silence compile warning on some BSDs and have a return */
 516                 assert(0);
 517                 return -1;
 518         }
 519 }
 520
 521 static int compare_block_infos(const void *bs1, const void *bs2)
 522 {
 523         uint32_t block1 = *(uint32_t *)bs1;
 524         uint32_t block2 = *(uint32_t *)bs2;
 525         int state1 = BLOCK_INFO_STATE(block1);
 526         int state2 = BLOCK_INFO_STATE(block2);
 527         int bscat1 = block_state_category(state1);
 528         int bscat2 = block_state_category(state2);
 529         int cycles1 = BLOCK_INFO_TRIMS(block1);
 530         int cycles2 = BLOCK_INFO_TRIMS(block2);
 531
 532         if (bscat1 < bscat2)
 533                 return -1;
 534         if (bscat1 > bscat2)
 535                 return 1;
 536
 537         if (cycles1 < cycles2)
 538                 return -1;
 539         if (cycles1 > cycles2)
 540                 return 1;
 541
 542         if (state1 < state2)
 543                 return -1;
 544         if (state1 > state2)
 545                 return 1;
 546
 547         assert(block1 == block2);
 548         return 0;
 549 }
 550
 551 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 552                                   fio_fp64_t *plist, unsigned int **percentiles,
 553                                   unsigned int *types)
 554 {
 555         int len = 0;
 556         int i, nr_uninit;
 557
 558         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 559
 560         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 561                 len++;
 562
 563         if (!len)
 564                 return 0;
 565
 566         /*
 567          * Sort the percentile list. Note that it may already be sorted if
 568          * we are using the default values, but since it's a short list this
 569          * isn't a worry. Also note that this does not work for NaN values.
 570          */
 571         if (len > 1)
 572                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 573
 574         nr_uninit = 0;
 575         /* Start only after the uninit entries end */
 576         for (nr_uninit = 0;
 577              nr_uninit < nr_block_infos
 578                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 579              nr_uninit ++)
 580                 ;
 581
 582         if (nr_uninit == nr_block_infos)
 583                 return 0;
 584
 585         *percentiles = calloc(len, sizeof(**percentiles));
 586
 587         for (i = 0; i < len; i++) {
 588                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 589                                 + nr_uninit;
 590                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 591         }
 592
 593         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 594         for (i = 0; i < nr_block_infos; i++)
 595                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 596
 597         return len;
 598 }
 599
 600 static const char *block_state_names[] = {
 601         [BLOCK_STATE_UNINIT] = "unwritten",
 602         [BLOCK_STATE_TRIMMED] = "trimmed",
 603         [BLOCK_STATE_WRITTEN] = "written",
 604         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 605         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 606 };
 607
 608 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 609                              fio_fp64_t *plist, struct buf_output *out)
 610 {
 611         int len, pos, i;
 612         unsigned int *percentiles = NULL;
 613         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 614
 615         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 616                                      &percentiles, block_state_counts);
 617
 618         log_buf(out, "  block lifetime percentiles :\n   |");
 619         pos = 0;
 620         for (i = 0; i < len; i++) {
 621                 uint32_t block_info = percentiles[i];
 622 #define LINE_LENGTH     75
 623                 char str[LINE_LENGTH];
 624                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 625                                      plist[i].u.f, block_info,
 626                                      i == len - 1 ? '\n' : ',');
 627                 assert(strln < LINE_LENGTH);
 628                 if (pos + strln > LINE_LENGTH) {
 629                         pos = 0;
 630                         log_buf(out, "\n   |");
 631                 }
 632                 log_buf(out, "%s", str);
 633                 pos += strln;
 634 #undef LINE_LENGTH
 635         }
 636         if (percentiles)
 637                 free(percentiles);
 638
 639         log_buf(out, "        states               :");
 640         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 641                 log_buf(out, " %s=%u%c",
 642                          block_state_names[i], block_state_counts[i],
 643                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 644 }
 645
 646 static void show_thread_status_normal(struct thread_stat *ts,
 647                                       struct group_run_stats *rs,
 648                                       struct buf_output *out)
 649 {
 650         double usr_cpu, sys_cpu;
 651         unsigned long runtime;
 652         double io_u_dist[FIO_IO_U_MAP_NR];
 653         time_t time_p;
 654         char time_buf[32];
 655
 656         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 657                 return;
 658
 659         time(&time_p);
 660         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 661
 662         if (!ts->error) {
 663                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 664                                         ts->name, ts->groupid, ts->members,
 665                                         ts->error, (int) ts->pid, time_buf);
 666         } else {
 667                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 668                                         ts->name, ts->groupid, ts->members,
 669                                         ts->error, ts->verror, (int) ts->pid,
 670                                         time_buf);
 671         }
 672
 673         if (strlen(ts->description))
 674                 log_buf(out, "  Description  : [%s]\n", ts->description);
 675
 676         if (ts->io_bytes[DDIR_READ])
 677                 show_ddir_status(rs, ts, DDIR_READ, out);
 678         if (ts->io_bytes[DDIR_WRITE])
 679                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 680         if (ts->io_bytes[DDIR_TRIM])
 681                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 682
 683         show_latencies(ts, out);
 684
 685         runtime = ts->total_run_time;
 686         if (runtime) {
 687                 double runt = (double) runtime;
 688
 689                 usr_cpu = (double) ts->usr_time * 100 / runt;
 690                 sys_cpu = (double) ts->sys_time * 100 / runt;
 691         } else {
 692                 usr_cpu = 0;
 693                 sys_cpu = 0;
 694         }
 695
 696         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 697                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 698                         (unsigned long long) ts->ctx,
 699                         (unsigned long long) ts->majf,
 700                         (unsigned long long) ts->minf);
 701
 702         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 703         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 704                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 705                                         io_u_dist[1], io_u_dist[2],
 706                                         io_u_dist[3], io_u_dist[4],
 707                                         io_u_dist[5], io_u_dist[6]);
 708
 709         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 710         log_buf(out, "     submit    : 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         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 716         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 717                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 718                                         io_u_dist[1], io_u_dist[2],
 719                                         io_u_dist[3], io_u_dist[4],
 720                                         io_u_dist[5], io_u_dist[6]);
 721         log_buf(out, "     issued    : total=r=%llu/w=%llu/d=%llu,"
 722                                  " short=r=%llu/w=%llu/d=%llu,"
 723                                  " drop=r=%llu/w=%llu/d=%llu\n",
 724                                         (unsigned long long) ts->total_io_u[0],
 725                                         (unsigned long long) ts->total_io_u[1],
 726                                         (unsigned long long) ts->total_io_u[2],
 727                                         (unsigned long long) ts->short_io_u[0],
 728                                         (unsigned long long) ts->short_io_u[1],
 729                                         (unsigned long long) ts->short_io_u[2],
 730                                         (unsigned long long) ts->drop_io_u[0],
 731                                         (unsigned long long) ts->drop_io_u[1],
 732                                         (unsigned long long) ts->drop_io_u[2]);
 733         if (ts->continue_on_error) {
 734                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 735                                         (unsigned long long)ts->total_err_count,
 736                                         ts->first_error,
 737                                         strerror(ts->first_error));
 738         }
 739         if (ts->latency_depth) {
 740                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 741                                         (unsigned long long)ts->latency_target,
 742                                         (unsigned long long)ts->latency_window,
 743                                         ts->latency_percentile.u.f,
 744                                         ts->latency_depth);
 745         }
 746
 747         if (ts->nr_block_infos)
 748                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 749                                   ts->percentile_list, out);
 750 }
 751
 752 static void show_ddir_status_terse(struct thread_stat *ts,
 753                                    struct group_run_stats *rs, int ddir,
 754                                    struct buf_output *out)
 755 {
 756         unsigned long min, max;
 757         unsigned long long bw, iops;
 758         unsigned int *ovals = NULL;
 759         double mean, dev;
 760         unsigned int len, minv, maxv;
 761         int i;
 762
 763         assert(ddir_rw(ddir));
 764
 765         iops = bw = 0;
 766         if (ts->runtime[ddir]) {
 767                 uint64_t runt = ts->runtime[ddir];
 768
 769                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 770                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 771         }
 772
 773         log_buf(out, ";%llu;%llu;%llu;%llu",
 774                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 775                                         (unsigned long long) ts->runtime[ddir]);
 776
 777         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 778                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 779         else
 780                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 781
 782         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 783                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 784         else
 785                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 786
 787         if (ts->clat_percentiles) {
 788                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 789                                         ts->clat_stat[ddir].samples,
 790                                         ts->percentile_list, &ovals, &maxv,
 791                                         &minv);
 792         } else
 793                 len = 0;
 794
 795         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 796                 if (i >= len) {
 797                         log_buf(out, ";0%%=0");
 798                         continue;
 799                 }
 800                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 801         }
 802
 803         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 804                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 805         else
 806                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 807
 808         if (ovals)
 809                 free(ovals);
 810
 811         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 812                 double p_of_agg = 100.0;
 813
 814                 if (rs->agg[ddir]) {
 815                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 816                         if (p_of_agg > 100.0)
 817                                 p_of_agg = 100.0;
 818                 }
 819
 820                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 821         } else
 822                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 823 }
 824
 825 static void add_ddir_status_json(struct thread_stat *ts,
 826                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 827 {
 828         unsigned long min, max;
 829         unsigned long long bw;
 830         unsigned int *ovals = NULL;
 831         double mean, dev, iops;
 832         unsigned int len, minv, maxv;
 833         int i;
 834         const char *ddirname[] = {"read", "write", "trim"};
 835         struct json_object *dir_object, *tmp_object, *percentile_object;
 836         char buf[120];
 837         double p_of_agg = 100.0;
 838
 839         assert(ddir_rw(ddir));
 840
 841         if (ts->unified_rw_rep && ddir != DDIR_READ)
 842                 return;
 843
 844         dir_object = json_create_object();
 845         json_object_add_value_object(parent,
 846                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 847
 848         bw = 0;
 849         iops = 0.0;
 850         if (ts->runtime[ddir]) {
 851                 uint64_t runt = ts->runtime[ddir];
 852
 853                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 854                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 855         }
 856
 857         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 858         json_object_add_value_int(dir_object, "bw", bw);
 859         json_object_add_value_float(dir_object, "iops", iops);
 860         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 861         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 862         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 863         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 864
 865         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 866                 min = max = 0;
 867                 mean = dev = 0.0;
 868         }
 869         tmp_object = json_create_object();
 870         json_object_add_value_object(dir_object, "slat", tmp_object);
 871         json_object_add_value_int(tmp_object, "min", min);
 872         json_object_add_value_int(tmp_object, "max", max);
 873         json_object_add_value_float(tmp_object, "mean", mean);
 874         json_object_add_value_float(tmp_object, "stddev", dev);
 875
 876         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 877                 min = max = 0;
 878                 mean = dev = 0.0;
 879         }
 880         tmp_object = json_create_object();
 881         json_object_add_value_object(dir_object, "clat", tmp_object);
 882         json_object_add_value_int(tmp_object, "min", min);
 883         json_object_add_value_int(tmp_object, "max", max);
 884         json_object_add_value_float(tmp_object, "mean", mean);
 885         json_object_add_value_float(tmp_object, "stddev", dev);
 886
 887         if (ts->clat_percentiles) {
 888                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 889                                         ts->clat_stat[ddir].samples,
 890                                         ts->percentile_list, &ovals, &maxv,
 891                                         &minv);
 892         } else
 893                 len = 0;
 894
 895         percentile_object = json_create_object();
 896         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 897         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 898                 if (i >= len) {
 899                         json_object_add_value_int(percentile_object, "0.00", 0);
 900                         continue;
 901                 }
 902                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 903                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 904         }
 905
 906         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 907                 min = max = 0;
 908                 mean = dev = 0.0;
 909         }
 910         tmp_object = json_create_object();
 911         json_object_add_value_object(dir_object, "lat", tmp_object);
 912         json_object_add_value_int(tmp_object, "min", min);
 913         json_object_add_value_int(tmp_object, "max", max);
 914         json_object_add_value_float(tmp_object, "mean", mean);
 915         json_object_add_value_float(tmp_object, "stddev", dev);
 916         if (ovals)
 917                 free(ovals);
 918
 919         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 920                 if (rs->agg[ddir]) {
 921                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 922                         if (p_of_agg > 100.0)
 923                                 p_of_agg = 100.0;
 924                 }
 925         } else {
 926                 min = max = 0;
 927                 p_of_agg = mean = dev = 0.0;
 928         }
 929         json_object_add_value_int(dir_object, "bw_min", min);
 930         json_object_add_value_int(dir_object, "bw_max", max);
 931         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 932         json_object_add_value_float(dir_object, "bw_mean", mean);
 933         json_object_add_value_float(dir_object, "bw_dev", dev);
 934 }
 935
 936 static void show_thread_status_terse_v2(struct thread_stat *ts,
 937                                         struct group_run_stats *rs,
 938                                         struct buf_output *out)
 939 {
 940         double io_u_dist[FIO_IO_U_MAP_NR];
 941         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 942         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 943         double usr_cpu, sys_cpu;
 944         int i;
 945
 946         /* General Info */
 947         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 948         /* Log Read Status */
 949         show_ddir_status_terse(ts, rs, DDIR_READ, out);
 950         /* Log Write Status */
 951         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
 952         /* Log Trim Status */
 953         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
 954
 955         /* CPU Usage */
 956         if (ts->total_run_time) {
 957                 double runt = (double) ts->total_run_time;
 958
 959                 usr_cpu = (double) ts->usr_time * 100 / runt;
 960                 sys_cpu = (double) ts->sys_time * 100 / runt;
 961         } else {
 962                 usr_cpu = 0;
 963                 sys_cpu = 0;
 964         }
 965
 966         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 967                                                 (unsigned long long) ts->ctx,
 968                                                 (unsigned long long) ts->majf,
 969                                                 (unsigned long long) ts->minf);
 970
 971         /* Calc % distribution of IO depths, usecond, msecond latency */
 972         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 973         stat_calc_lat_u(ts, io_u_lat_u);
 974         stat_calc_lat_m(ts, io_u_lat_m);
 975
 976         /* Only show fixed 7 I/O depth levels*/
 977         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 978                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 979                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 980
 981         /* Microsecond latency */
 982         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 983                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
 984         /* Millisecond latency */
 985         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 986                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
 987         /* Additional output if continue_on_error set - default off*/
 988         if (ts->continue_on_error)
 989                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
 990         log_buf(out, "\n");
 991
 992         /* Additional output if description is set */
 993         if (strlen(ts->description))
 994                 log_buf(out, ";%s", ts->description);
 995
 996         log_buf(out, "\n");
 997 }
 998
 999 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1000                                            struct group_run_stats *rs, int ver,
1001                                            struct buf_output *out)
1002 {
1003         double io_u_dist[FIO_IO_U_MAP_NR];
1004         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1005         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1006         double usr_cpu, sys_cpu;
1007         int i;
1008
1009         /* General Info */
1010         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1011                                         ts->name, ts->groupid, ts->error);
1012         /* Log Read Status */
1013         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1014         /* Log Write Status */
1015         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1016         /* Log Trim Status */
1017         if (ver == 4)
1018                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1019
1020         /* CPU Usage */
1021         if (ts->total_run_time) {
1022                 double runt = (double) ts->total_run_time;
1023
1024                 usr_cpu = (double) ts->usr_time * 100 / runt;
1025                 sys_cpu = (double) ts->sys_time * 100 / runt;
1026         } else {
1027                 usr_cpu = 0;
1028                 sys_cpu = 0;
1029         }
1030
1031         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1032                                                 (unsigned long long) ts->ctx,
1033                                                 (unsigned long long) ts->majf,
1034                                                 (unsigned long long) ts->minf);
1035
1036         /* Calc % distribution of IO depths, usecond, msecond latency */
1037         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1038         stat_calc_lat_u(ts, io_u_lat_u);
1039         stat_calc_lat_m(ts, io_u_lat_m);
1040
1041         /* Only show fixed 7 I/O depth levels*/
1042         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1043                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1044                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1045
1046         /* Microsecond latency */
1047         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1048                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1049         /* Millisecond latency */
1050         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1051                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1052
1053         /* disk util stats, if any */
1054         show_disk_util(1, NULL, out);
1055
1056         /* Additional output if continue_on_error set - default off*/
1057         if (ts->continue_on_error)
1058                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1059
1060         /* Additional output if description is set */
1061         if (strlen(ts->description))
1062                 log_buf(out, ";%s", ts->description);
1063
1064         log_buf(out, "\n");
1065 }
1066
1067 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1068                                     struct group_run_stats *rs)
1069 {
1070         struct json_object *root, *tmp;
1071         struct jobs_eta *je;
1072         double io_u_dist[FIO_IO_U_MAP_NR];
1073         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1074         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1075         double usr_cpu, sys_cpu;
1076         int i;
1077         size_t size;
1078
1079         root = json_create_object();
1080         json_object_add_value_string(root, "jobname", ts->name);
1081         json_object_add_value_int(root, "groupid", ts->groupid);
1082         json_object_add_value_int(root, "error", ts->error);
1083
1084         /* ETA Info */
1085         je = get_jobs_eta(1, &size);
1086         json_object_add_value_int(root, "eta", je->eta_sec);
1087         json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1088
1089
1090         add_ddir_status_json(ts, rs, DDIR_READ, root);
1091         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1092         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1093
1094         /* CPU Usage */
1095         if (ts->total_run_time) {
1096                 double runt = (double) ts->total_run_time;
1097
1098                 usr_cpu = (double) ts->usr_time * 100 / runt;
1099                 sys_cpu = (double) ts->sys_time * 100 / runt;
1100         } else {
1101                 usr_cpu = 0;
1102                 sys_cpu = 0;
1103         }
1104         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1105         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1106         json_object_add_value_int(root, "ctx", ts->ctx);
1107         json_object_add_value_int(root, "majf", ts->majf);
1108         json_object_add_value_int(root, "minf", ts->minf);
1109
1110
1111         /* Calc % distribution of IO depths, usecond, msecond latency */
1112         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1113         stat_calc_lat_u(ts, io_u_lat_u);
1114         stat_calc_lat_m(ts, io_u_lat_m);
1115
1116         tmp = json_create_object();
1117         json_object_add_value_object(root, "iodepth_level", tmp);
1118         /* Only show fixed 7 I/O depth levels*/
1119         for (i = 0; i < 7; i++) {
1120                 char name[20];
1121                 if (i < 6)
1122                         snprintf(name, 20, "%d", 1 << i);
1123                 else
1124                         snprintf(name, 20, ">=%d", 1 << i);
1125                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1126         }
1127
1128         tmp = json_create_object();
1129         json_object_add_value_object(root, "latency_us", tmp);
1130         /* Microsecond latency */
1131         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1132                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1133                                  "250", "500", "750", "1000", };
1134                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1135         }
1136         /* Millisecond latency */
1137         tmp = json_create_object();
1138         json_object_add_value_object(root, "latency_ms", tmp);
1139         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1140                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1141                                  "250", "500", "750", "1000", "2000",
1142                                  ">=2000", };
1143                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1144         }
1145
1146         /* Additional output if continue_on_error set - default off*/
1147         if (ts->continue_on_error) {
1148                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1149                 json_object_add_value_int(root, "first_error", ts->first_error);
1150         }
1151
1152         if (ts->latency_depth) {
1153                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1154                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1155                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1156                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1157         }
1158
1159         /* Additional output if description is set */
1160         if (strlen(ts->description))
1161                 json_object_add_value_string(root, "desc", ts->description);
1162
1163         if (ts->nr_block_infos) {
1164                 /* Block error histogram and types */
1165                 int len;
1166                 unsigned int *percentiles = NULL;
1167                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1168
1169                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1170                                              ts->percentile_list,
1171                                              &percentiles, block_state_counts);
1172
1173                 if (len) {
1174                         struct json_object *block, *percentile_object, *states;
1175                         int state, i;
1176                         block = json_create_object();
1177                         json_object_add_value_object(root, "block", block);
1178
1179                         percentile_object = json_create_object();
1180                         json_object_add_value_object(block, "percentiles",
1181                                                      percentile_object);
1182                         for (i = 0; i < len; i++) {
1183                                 char buf[20];
1184                                 snprintf(buf, sizeof(buf), "%f",
1185                                          ts->percentile_list[i].u.f);
1186                                 json_object_add_value_int(percentile_object,
1187                                                           (const char *)buf,
1188                                                           percentiles[i]);
1189                         }
1190
1191                         states = json_create_object();
1192                         json_object_add_value_object(block, "states", states);
1193                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1194                                 json_object_add_value_int(states,
1195                                         block_state_names[state],
1196                                         block_state_counts[state]);
1197                         }
1198                         free(percentiles);
1199                 }
1200         }
1201
1202         return root;
1203 }
1204
1205 static void show_thread_status_terse(struct thread_stat *ts,
1206                                      struct group_run_stats *rs,
1207                                      struct buf_output *out)
1208 {
1209         if (terse_version == 2)
1210                 show_thread_status_terse_v2(ts, rs, out);
1211         else if (terse_version == 3 || terse_version == 4)
1212                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1213         else
1214                 log_err("fio: bad terse version!? %d\n", terse_version);
1215 }
1216
1217 struct json_object *show_thread_status(struct thread_stat *ts,
1218                                        struct group_run_stats *rs,
1219                                        struct buf_output *out)
1220 {
1221         struct json_object *ret = NULL;
1222
1223         if (output_format & FIO_OUTPUT_TERSE)
1224                 show_thread_status_terse(ts, rs,  out);
1225         if (output_format & FIO_OUTPUT_JSON)
1226                 ret = show_thread_status_json(ts, rs);
1227         if (output_format & FIO_OUTPUT_NORMAL)
1228                 show_thread_status_normal(ts, rs,  out);
1229
1230         return ret;
1231 }
1232
1233 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1234 {
1235         double mean, S;
1236
1237         if (src->samples == 0)
1238                 return;
1239
1240         dst->min_val = min(dst->min_val, src->min_val);
1241         dst->max_val = max(dst->max_val, src->max_val);
1242
1243         /*
1244          * Compute new mean and S after the merge
1245          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1246          *  #Parallel_algorithm>
1247          */
1248         if (nr == 1) {
1249                 mean = src->mean.u.f;
1250                 S = src->S.u.f;
1251         } else {
1252                 double delta = src->mean.u.f - dst->mean.u.f;
1253
1254                 mean = ((src->mean.u.f * src->samples) +
1255                         (dst->mean.u.f * dst->samples)) /
1256                         (dst->samples + src->samples);
1257
1258                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1259                         (dst->samples * src->samples) /
1260                         (dst->samples + src->samples);
1261         }
1262
1263         dst->samples += src->samples;
1264         dst->mean.u.f = mean;
1265         dst->S.u.f = S;
1266 }
1267
1268 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1269 {
1270         int i;
1271
1272         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1273                 if (dst->max_run[i] < src->max_run[i])
1274                         dst->max_run[i] = src->max_run[i];
1275                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1276                         dst->min_run[i] = src->min_run[i];
1277                 if (dst->max_bw[i] < src->max_bw[i])
1278                         dst->max_bw[i] = src->max_bw[i];
1279                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1280                         dst->min_bw[i] = src->min_bw[i];
1281
1282                 dst->io_kb[i] += src->io_kb[i];
1283                 dst->agg[i] += src->agg[i];
1284         }
1285
1286         if (!dst->kb_base)
1287                 dst->kb_base = src->kb_base;
1288         if (!dst->unit_base)
1289                 dst->unit_base = src->unit_base;
1290 }
1291
1292 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1293 {
1294         int l, k;
1295
1296         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1297                 if (!dst->unified_rw_rep) {
1298                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1299                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1300                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1301                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1302
1303                         dst->io_bytes[l] += src->io_bytes[l];
1304
1305                         if (dst->runtime[l] < src->runtime[l])
1306                                 dst->runtime[l] = src->runtime[l];
1307                 } else {
1308                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1309                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1310                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1311                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1312
1313                         dst->io_bytes[0] += src->io_bytes[l];
1314
1315                         if (dst->runtime[0] < src->runtime[l])
1316                                 dst->runtime[0] = src->runtime[l];
1317                 }
1318         }
1319
1320         dst->usr_time += src->usr_time;
1321         dst->sys_time += src->sys_time;
1322         dst->ctx += src->ctx;
1323         dst->majf += src->majf;
1324         dst->minf += src->minf;
1325
1326         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1327                 dst->io_u_map[k] += src->io_u_map[k];
1328         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1329                 dst->io_u_submit[k] += src->io_u_submit[k];
1330         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1331                 dst->io_u_complete[k] += src->io_u_complete[k];
1332         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1333                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1334         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1335                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1336
1337         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1338                 if (!dst->unified_rw_rep) {
1339                         dst->total_io_u[k] += src->total_io_u[k];
1340                         dst->short_io_u[k] += src->short_io_u[k];
1341                         dst->drop_io_u[k] += src->drop_io_u[k];
1342                 } else {
1343                         dst->total_io_u[0] += src->total_io_u[k];
1344                         dst->short_io_u[0] += src->short_io_u[k];
1345                         dst->drop_io_u[0] += src->drop_io_u[k];
1346                 }
1347         }
1348
1349         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1350                 int m;
1351
1352                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1353                         if (!dst->unified_rw_rep)
1354                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1355                         else
1356                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1357                 }
1358         }
1359
1360         dst->total_run_time += src->total_run_time;
1361         dst->total_submit += src->total_submit;
1362         dst->total_complete += src->total_complete;
1363 }
1364
1365 void init_group_run_stat(struct group_run_stats *gs)
1366 {
1367         int i;
1368         memset(gs, 0, sizeof(*gs));
1369
1370         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1371                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1372 }
1373
1374 void init_thread_stat(struct thread_stat *ts)
1375 {
1376         int j;
1377
1378         memset(ts, 0, sizeof(*ts));
1379
1380         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1381                 ts->lat_stat[j].min_val = -1UL;
1382                 ts->clat_stat[j].min_val = -1UL;
1383                 ts->slat_stat[j].min_val = -1UL;
1384                 ts->bw_stat[j].min_val = -1UL;
1385         }
1386         ts->groupid = -1;
1387 }
1388
1389 void __show_run_stats(void)
1390 {
1391         struct group_run_stats *runstats, *rs;
1392         struct thread_data *td;
1393         struct thread_stat *threadstats, *ts;
1394         int i, j, k, nr_ts, last_ts, idx;
1395         int kb_base_warned = 0;
1396         int unit_base_warned = 0;
1397         struct json_object *root = NULL;
1398         struct json_array *array = NULL;
1399         struct buf_output output[FIO_OUTPUT_NR];
1400
1401         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1402
1403         memset(output, 0, sizeof(output));
1404
1405         for (i = 0; i < groupid + 1; i++)
1406                 init_group_run_stat(&runstats[i]);
1407
1408         /*
1409          * find out how many threads stats we need. if group reporting isn't
1410          * enabled, it's one-per-td.
1411          */
1412         nr_ts = 0;
1413         last_ts = -1;
1414         for_each_td(td, i) {
1415                 if (!td->o.group_reporting) {
1416                         nr_ts++;
1417                         continue;
1418                 }
1419                 if (last_ts == td->groupid)
1420                         continue;
1421
1422                 last_ts = td->groupid;
1423                 nr_ts++;
1424         }
1425
1426         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1427
1428         for (i = 0; i < nr_ts; i++)
1429                 init_thread_stat(&threadstats[i]);
1430
1431         j = 0;
1432         last_ts = -1;
1433         idx = 0;
1434         for_each_td(td, i) {
1435                 if (idx && (!td->o.group_reporting ||
1436                     (td->o.group_reporting && last_ts != td->groupid))) {
1437                         idx = 0;
1438                         j++;
1439                 }
1440
1441                 last_ts = td->groupid;
1442
1443                 ts = &threadstats[j];
1444
1445                 ts->clat_percentiles = td->o.clat_percentiles;
1446                 ts->percentile_precision = td->o.percentile_precision;
1447                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1448
1449                 idx++;
1450                 ts->members++;
1451
1452                 if (ts->groupid == -1) {
1453                         /*
1454                          * These are per-group shared already
1455                          */
1456                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1457                         if (td->o.description)
1458                                 strncpy(ts->description, td->o.description,
1459                                                 FIO_JOBDESC_SIZE - 1);
1460                         else
1461                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1462
1463                         /*
1464                          * If multiple entries in this group, this is
1465                          * the first member.
1466                          */
1467                         ts->thread_number = td->thread_number;
1468                         ts->groupid = td->groupid;
1469
1470                         /*
1471                          * first pid in group, not very useful...
1472                          */
1473                         ts->pid = td->pid;
1474
1475                         ts->kb_base = td->o.kb_base;
1476                         ts->unit_base = td->o.unit_base;
1477                         ts->unified_rw_rep = td->o.unified_rw_rep;
1478                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1479                         log_info("fio: kb_base differs for jobs in group, using"
1480                                  " %u as the base\n", ts->kb_base);
1481                         kb_base_warned = 1;
1482                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1483                         log_info("fio: unit_base differs for jobs in group, using"
1484                                  " %u as the base\n", ts->unit_base);
1485                         unit_base_warned = 1;
1486                 }
1487
1488                 ts->continue_on_error = td->o.continue_on_error;
1489                 ts->total_err_count += td->total_err_count;
1490                 ts->first_error = td->first_error;
1491                 if (!ts->error) {
1492                         if (!td->error && td->o.continue_on_error &&
1493                             td->first_error) {
1494                                 ts->error = td->first_error;
1495                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1496                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1497                         } else  if (td->error) {
1498                                 ts->error = td->error;
1499                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1500                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1501                         }
1502                 }
1503
1504                 ts->latency_depth = td->latency_qd;
1505                 ts->latency_target = td->o.latency_target;
1506                 ts->latency_percentile = td->o.latency_percentile;
1507                 ts->latency_window = td->o.latency_window;
1508
1509                 ts->nr_block_infos = td->ts.nr_block_infos;
1510                 for (k = 0; k < ts->nr_block_infos; k++)
1511                         ts->block_infos[k] = td->ts.block_infos[k];
1512
1513                 sum_thread_stats(ts, &td->ts, idx);
1514         }
1515
1516         for (i = 0; i < nr_ts; i++) {
1517                 unsigned long long bw;
1518
1519                 ts = &threadstats[i];
1520                 rs = &runstats[ts->groupid];
1521                 rs->kb_base = ts->kb_base;
1522                 rs->unit_base = ts->unit_base;
1523                 rs->unified_rw_rep += ts->unified_rw_rep;
1524
1525                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1526                         if (!ts->runtime[j])
1527                                 continue;
1528                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1529                                 rs->min_run[j] = ts->runtime[j];
1530                         if (ts->runtime[j] > rs->max_run[j])
1531                                 rs->max_run[j] = ts->runtime[j];
1532
1533                         bw = 0;
1534                         if (ts->runtime[j]) {
1535                                 unsigned long runt = ts->runtime[j];
1536                                 unsigned long long kb;
1537
1538                                 kb = ts->io_bytes[j] / rs->kb_base;
1539                                 bw = kb * 1000 / runt;
1540                         }
1541                         if (bw < rs->min_bw[j])
1542                                 rs->min_bw[j] = bw;
1543                         if (bw > rs->max_bw[j])
1544                                 rs->max_bw[j] = bw;
1545
1546                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1547                 }
1548         }
1549
1550         for (i = 0; i < groupid + 1; i++) {
1551                 int ddir;
1552
1553                 rs = &runstats[i];
1554
1555                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1556                         if (rs->max_run[ddir])
1557                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1558                                                 rs->max_run[ddir];
1559                 }
1560         }
1561
1562         for (i = 0; i < FIO_OUTPUT_NR; i++)
1563                 buf_output_init(&output[i], i);
1564
1565         /*
1566          * don't overwrite last signal output
1567          */
1568         if (output_format & FIO_OUTPUT_NORMAL)
1569                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1570         if (output_format & FIO_OUTPUT_JSON) {
1571                 char time_buf[32];
1572                 time_t time_p;
1573
1574                 time(&time_p);
1575                 os_ctime_r((const time_t *) &time_p, time_buf,
1576                                 sizeof(time_buf));
1577                 time_buf[strlen(time_buf) - 1] = '\0';
1578
1579                 root = json_create_object();
1580                 json_object_add_value_string(root, "fio version", fio_version_string);
1581                 json_object_add_value_int(root, "timestamp", time_p);
1582                 json_object_add_value_string(root, "time", time_buf);
1583                 array = json_create_array();
1584                 json_object_add_value_array(root, "jobs", array);
1585         }
1586
1587         for (i = 0; i < nr_ts; i++) {
1588                 ts = &threadstats[i];
1589                 rs = &runstats[ts->groupid];
1590
1591                 if (is_backend)
1592                         fio_server_send_ts(ts, rs);
1593                 else {
1594                         if (output_format & FIO_OUTPUT_TERSE)
1595                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1596                         if (output_format & FIO_OUTPUT_JSON) {
1597                                 struct json_object *tmp = show_thread_status_json(ts, rs);
1598                                 json_array_add_value_object(array, tmp);
1599                         }
1600                         if (output_format & FIO_OUTPUT_NORMAL)
1601                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1602                 }
1603         }
1604         if (output_format & FIO_OUTPUT_JSON) {
1605                 /* disk util stats, if any */
1606                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1607
1608                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1609
1610                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1611                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1612                 json_free_object(root);
1613         }
1614
1615         for (i = 0; i < groupid + 1; i++) {
1616                 rs = &runstats[i];
1617
1618                 rs->groupid = i;
1619                 if (is_backend)
1620                         fio_server_send_gs(rs);
1621                 else if (output_format & FIO_OUTPUT_NORMAL)
1622                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1623         }
1624
1625         if (is_backend)
1626                 fio_server_send_du();
1627         else if (output_format & FIO_OUTPUT_NORMAL) {
1628                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1629                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1630         }
1631
1632         for (i = 0; i < FIO_OUTPUT_NR; i++)
1633                 buf_output_flush(&output[i]);
1634
1635         log_info_flush();
1636         free(runstats);
1637         free(threadstats);
1638
1639         for (i = 0; i < FIO_OUTPUT_NR; i++)
1640                 buf_output_free(&output[i]);
1641 }
1642
1643 void show_run_stats(void)
1644 {
1645         fio_mutex_down(stat_mutex);
1646         __show_run_stats();
1647         fio_mutex_up(stat_mutex);
1648 }
1649
1650 void __show_running_run_stats(void)
1651 {
1652         struct thread_data *td;
1653         unsigned long long *rt;
1654         struct timeval tv;
1655         int i;
1656
1657         fio_mutex_down(stat_mutex);
1658
1659         rt = malloc(thread_number * sizeof(unsigned long long));
1660         fio_gettime(&tv, NULL);
1661
1662         for_each_td(td, i) {
1663                 rt[i] = mtime_since(&td->start, &tv);
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                 td->update_rusage = 1;
1672                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1673                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1674                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1675                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1676         }
1677
1678         for_each_td(td, i) {
1679                 if (td->runstate >= TD_EXITED)
1680                         continue;
1681                 if (td->rusage_sem) {
1682                         td->update_rusage = 1;
1683                         fio_mutex_down(td->rusage_sem);
1684                 }
1685                 td->update_rusage = 0;
1686         }
1687
1688         __show_run_stats();
1689
1690         for_each_td(td, i) {
1691                 if (td_read(td) && td->io_bytes[DDIR_READ])
1692                         td->ts.runtime[DDIR_READ] -= rt[i];
1693                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1694                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1695                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1696                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1697         }
1698
1699         free(rt);
1700         fio_mutex_up(stat_mutex);
1701 }
1702
1703 static int status_interval_init;
1704 static struct timeval status_time;
1705 static int status_file_disabled;
1706
1707 #define FIO_STATUS_FILE         "fio-dump-status"
1708
1709 static int check_status_file(void)
1710 {
1711         struct stat sb;
1712         const char *temp_dir;
1713         char fio_status_file_path[PATH_MAX];
1714
1715         if (status_file_disabled)
1716                 return 0;
1717
1718         temp_dir = getenv("TMPDIR");
1719         if (temp_dir == NULL) {
1720                 temp_dir = getenv("TEMP");
1721                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1722                         temp_dir = NULL;
1723         }
1724         if (temp_dir == NULL)
1725                 temp_dir = "/tmp";
1726
1727         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1728
1729         if (stat(fio_status_file_path, &sb))
1730                 return 0;
1731
1732         if (unlink(fio_status_file_path) < 0) {
1733                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1734                                                         strerror(errno));
1735                 log_err("fio: disabling status file updates\n");
1736                 status_file_disabled = 1;
1737         }
1738
1739         return 1;
1740 }
1741
1742 void check_for_running_stats(void)
1743 {
1744         if (status_interval) {
1745                 if (!status_interval_init) {
1746                         fio_gettime(&status_time, NULL);
1747                         status_interval_init = 1;
1748                 } else if (mtime_since_now(&status_time) >= status_interval) {
1749                         show_running_run_stats();
1750                         fio_gettime(&status_time, NULL);
1751                         return;
1752                 }
1753         }
1754         if (check_status_file()) {
1755                 show_running_run_stats();
1756                 return;
1757         }
1758 }
1759
1760 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1761 {
1762         double val = data;
1763         double delta;
1764
1765         if (data > is->max_val)
1766                 is->max_val = data;
1767         if (data < is->min_val)
1768                 is->min_val = data;
1769
1770         delta = val - is->mean.u.f;
1771         if (delta) {
1772                 is->mean.u.f += delta / (is->samples + 1.0);
1773                 is->S.u.f += delta * (val - is->mean.u.f);
1774         }
1775
1776         is->samples++;
1777 }
1778
1779 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1780                              enum fio_ddir ddir, unsigned int bs,
1781                              unsigned long t, uint64_t offset)
1782 {
1783         uint64_t nr_samples = iolog->nr_samples;
1784         struct io_sample *s;
1785
1786         if (iolog->disabled)
1787                 return;
1788
1789         if (!iolog->nr_samples)
1790                 iolog->avg_last = t;
1791
1792         if (iolog->nr_samples == iolog->max_samples) {
1793                 size_t new_size;
1794                 void *new_log;
1795
1796                 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1797
1798                 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1799                         if (iolog_flush(iolog, 0)) {
1800                                 log_err("fio: failed flushing iolog! Will stop logging.\n");
1801                                 iolog->disabled = 1;
1802                                 return;
1803                         }
1804                         nr_samples = iolog->nr_samples;
1805                 } else {
1806                         new_log = realloc(iolog->log, new_size);
1807                         if (!new_log) {
1808                                 log_err("fio: failed extending iolog! Will stop logging.\n");
1809                                 iolog->disabled = 1;
1810                                 return;
1811                         }
1812                         iolog->log = new_log;
1813                         iolog->max_samples <<= 1;
1814                 }
1815         }
1816
1817         s = get_sample(iolog, nr_samples);
1818
1819         s->val = val;
1820         s->time = t;
1821         io_sample_set_ddir(iolog, s, ddir);
1822         s->bs = bs;
1823
1824         if (iolog->log_offset) {
1825                 struct io_sample_offset *so = (void *) s;
1826
1827                 so->offset = offset;
1828         }
1829
1830         iolog->nr_samples++;
1831 }
1832
1833 static inline void reset_io_stat(struct io_stat *ios)
1834 {
1835         ios->max_val = ios->min_val = ios->samples = 0;
1836         ios->mean.u.f = ios->S.u.f = 0;
1837 }
1838
1839 void reset_io_stats(struct thread_data *td)
1840 {
1841         struct thread_stat *ts = &td->ts;
1842         int i, j;
1843
1844         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1845                 reset_io_stat(&ts->clat_stat[i]);
1846                 reset_io_stat(&ts->slat_stat[i]);
1847                 reset_io_stat(&ts->lat_stat[i]);
1848                 reset_io_stat(&ts->bw_stat[i]);
1849                 reset_io_stat(&ts->iops_stat[i]);
1850
1851                 ts->io_bytes[i] = 0;
1852                 ts->runtime[i] = 0;
1853
1854                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1855                         ts->io_u_plat[i][j] = 0;
1856         }
1857
1858         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1859                 ts->io_u_map[i] = 0;
1860                 ts->io_u_submit[i] = 0;
1861                 ts->io_u_complete[i] = 0;
1862                 ts->io_u_lat_u[i] = 0;
1863                 ts->io_u_lat_m[i] = 0;
1864                 ts->total_submit = 0;
1865                 ts->total_complete = 0;
1866         }
1867
1868         for (i = 0; i < 3; i++) {
1869                 ts->total_io_u[i] = 0;
1870                 ts->short_io_u[i] = 0;
1871                 ts->drop_io_u[i] = 0;
1872         }
1873 }
1874
1875 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1876 {
1877         /*
1878          * Note an entry in the log. Use the mean from the logged samples,
1879          * making sure to properly round up. Only write a log entry if we
1880          * had actual samples done.
1881          */
1882         if (iolog->avg_window[DDIR_READ].samples) {
1883                 unsigned long mr;
1884
1885                 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1886                 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1887         }
1888         if (iolog->avg_window[DDIR_WRITE].samples) {
1889                 unsigned long mw;
1890
1891                 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1892                 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1893         }
1894         if (iolog->avg_window[DDIR_TRIM].samples) {
1895                 unsigned long mw;
1896
1897                 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1898                 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1899         }
1900
1901         reset_io_stat(&iolog->avg_window[DDIR_READ]);
1902         reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1903         reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1904 }
1905
1906 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1907                            unsigned long val, enum fio_ddir ddir,
1908                            unsigned int bs, uint64_t offset)
1909 {
1910         unsigned long elapsed, this_window;
1911
1912         if (!ddir_rw(ddir))
1913                 return;
1914
1915         elapsed = mtime_since_now(&td->epoch);
1916
1917         /*
1918          * If no time averaging, just add the log sample.
1919          */
1920         if (!iolog->avg_msec) {
1921                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1922                 return;
1923         }
1924
1925         /*
1926          * Add the sample. If the time period has passed, then
1927          * add that entry to the log and clear.
1928          */
1929         add_stat_sample(&iolog->avg_window[ddir], val);
1930
1931         /*
1932          * If period hasn't passed, adding the above sample is all we
1933          * need to do.
1934          */
1935         this_window = elapsed - iolog->avg_last;
1936         if (this_window < iolog->avg_msec)
1937                 return;
1938
1939         _add_stat_to_log(iolog, elapsed);
1940
1941         iolog->avg_last = elapsed;
1942 }
1943
1944 void finalize_logs(struct thread_data *td)
1945 {
1946         unsigned long elapsed;
1947
1948         elapsed = mtime_since_now(&td->epoch);
1949
1950         if (td->clat_log)
1951                 _add_stat_to_log(td->clat_log, elapsed);
1952         if (td->slat_log)
1953                 _add_stat_to_log(td->slat_log, elapsed);
1954         if (td->lat_log)
1955                 _add_stat_to_log(td->lat_log, elapsed);
1956         if (td->bw_log)
1957                 _add_stat_to_log(td->bw_log, elapsed);
1958         if (td->iops_log)
1959                 _add_stat_to_log(td->iops_log, elapsed);
1960 }
1961
1962 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1963 {
1964         struct io_log *iolog;
1965
1966         if (!ddir_rw(ddir))
1967                 return;
1968
1969         iolog = agg_io_log[ddir];
1970         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1971 }
1972
1973 static void add_clat_percentile_sample(struct thread_stat *ts,
1974                                 unsigned long usec, enum fio_ddir ddir)
1975 {
1976         unsigned int idx = plat_val_to_idx(usec);
1977         assert(idx < FIO_IO_U_PLAT_NR);
1978
1979         ts->io_u_plat[ddir][idx]++;
1980 }
1981
1982 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1983                      unsigned long usec, unsigned int bs, uint64_t offset)
1984 {
1985         struct thread_stat *ts = &td->ts;
1986
1987         if (!ddir_rw(ddir))
1988                 return;
1989
1990         td_io_u_lock(td);
1991
1992         add_stat_sample(&ts->clat_stat[ddir], usec);
1993
1994         if (td->clat_log)
1995                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1996
1997         if (ts->clat_percentiles)
1998                 add_clat_percentile_sample(ts, usec, ddir);
1999
2000         td_io_u_unlock(td);
2001 }
2002
2003 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2004                      unsigned long usec, unsigned int bs, uint64_t offset)
2005 {
2006         struct thread_stat *ts = &td->ts;
2007
2008         if (!ddir_rw(ddir))
2009                 return;
2010
2011         td_io_u_lock(td);
2012
2013         add_stat_sample(&ts->slat_stat[ddir], usec);
2014
2015         if (td->slat_log)
2016                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
2017
2018         td_io_u_unlock(td);
2019 }
2020
2021 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2022                     unsigned long usec, unsigned int bs, uint64_t offset)
2023 {
2024         struct thread_stat *ts = &td->ts;
2025
2026         if (!ddir_rw(ddir))
2027                 return;
2028
2029         td_io_u_lock(td);
2030
2031         add_stat_sample(&ts->lat_stat[ddir], usec);
2032
2033         if (td->lat_log)
2034                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
2035
2036         td_io_u_unlock(td);
2037 }
2038
2039 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2040                    struct timeval *t)
2041 {
2042         struct thread_stat *ts = &td->ts;
2043         unsigned long spent, rate;
2044
2045         if (!ddir_rw(ddir))
2046                 return;
2047
2048         spent = mtime_since(&td->bw_sample_time, t);
2049         if (spent < td->o.bw_avg_time)
2050                 return;
2051
2052         td_io_u_lock(td);
2053
2054         /*
2055          * Compute both read and write rates for the interval.
2056          */
2057         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2058                 uint64_t delta;
2059
2060                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2061                 if (!delta)
2062                         continue; /* No entries for interval */
2063
2064                 if (spent)
2065                         rate = delta * 1000 / spent / 1024;
2066                 else
2067                         rate = 0;
2068
2069                 add_stat_sample(&ts->bw_stat[ddir], rate);
2070
2071                 if (td->bw_log)
2072                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2073
2074                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2075         }
2076
2077         fio_gettime(&td->bw_sample_time, NULL);
2078         td_io_u_unlock(td);
2079 }
2080
2081 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2082                      struct timeval *t)
2083 {
2084         struct thread_stat *ts = &td->ts;
2085         unsigned long spent, iops;
2086
2087         if (!ddir_rw(ddir))
2088                 return;
2089
2090         spent = mtime_since(&td->iops_sample_time, t);
2091         if (spent < td->o.iops_avg_time)
2092                 return;
2093
2094         td_io_u_lock(td);
2095
2096         /*
2097          * Compute both read and write rates for the interval.
2098          */
2099         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2100                 uint64_t delta;
2101
2102                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2103                 if (!delta)
2104                         continue; /* No entries for interval */
2105
2106                 if (spent)
2107                         iops = (delta * 1000) / spent;
2108                 else
2109                         iops = 0;
2110
2111                 add_stat_sample(&ts->iops_stat[ddir], iops);
2112
2113                 if (td->iops_log)
2114                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2115
2116                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2117         }
2118
2119         fio_gettime(&td->iops_sample_time, NULL);
2120         td_io_u_unlock(td);
2121 }
2122
2123 void stat_init(void)
2124 {
2125         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2126 }
2127
2128 void stat_exit(void)
2129 {
2130         /*
2131          * When we have the mutex, we know out-of-band access to it
2132          * have ended.
2133          */
2134         fio_mutex_down(stat_mutex);
2135         fio_mutex_remove(stat_mutex);
2136 }
2137
2138 /*
2139  * Called from signal handler. Wake up status thread.
2140  */
2141 void show_running_run_stats(void)
2142 {
2143         helper_do_stat = 1;
2144         pthread_cond_signal(&helper_cond);
2145 }
2146
2147 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2148 {
2149         /* Ignore io_u's which span multiple blocks--they will just get
2150          * inaccurate counts. */
2151         int idx = (io_u->offset - io_u->file->file_offset)
2152                         / td->o.bs[DDIR_TRIM];
2153         uint32_t *info = &td->ts.block_infos[idx];
2154         assert(idx < td->ts.nr_block_infos);
2155         return info;
2156 }