stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17
  18 struct fio_mutex *stat_mutex;
  19
  20 void update_rusage_stat(struct thread_data *td)
  21 {
  22         struct thread_stat *ts = &td->ts;
  23
  24         fio_getrusage(&td->ru_end);
  25         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  26                                         &td->ru_end.ru_utime);
  27         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  28                                         &td->ru_end.ru_stime);
  29         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  30                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  31         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  32         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  33
  34         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  35 }
  36
  37 /*
  38  * Given a latency, return the index of the corresponding bucket in
  39  * the structure tracking percentiles.
  40  *
  41  * (1) find the group (and error bits) that the value (latency)
  42  * belongs to by looking at its MSB. (2) find the bucket number in the
  43  * group by looking at the index bits.
  44  *
  45  */
  46 static unsigned int plat_val_to_idx(unsigned int val)
  47 {
  48         unsigned int msb, error_bits, base, offset, idx;
  49
  50         /* Find MSB starting from bit 0 */
  51         if (val == 0)
  52                 msb = 0;
  53         else
  54                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  55
  56         /*
  57          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  58          * all bits of the sample as index
  59          */
  60         if (msb <= FIO_IO_U_PLAT_BITS)
  61                 return val;
  62
  63         /* Compute the number of error bits to discard*/
  64         error_bits = msb - FIO_IO_U_PLAT_BITS;
  65
  66         /* Compute the number of buckets before the group */
  67         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  68
  69         /*
  70          * Discard the error bits and apply the mask to find the
  71          * index for the buckets in the group
  72          */
  73         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  74
  75         /* Make sure the index does not exceed (array size - 1) */
  76         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  77                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  78
  79         return idx;
  80 }
  81
  82 /*
  83  * Convert the given index of the bucket array to the value
  84  * represented by the bucket
  85  */
  86 static unsigned int plat_idx_to_val(unsigned int idx)
  87 {
  88         unsigned int error_bits, k, base;
  89
  90         assert(idx < FIO_IO_U_PLAT_NR);
  91
  92         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  93          * all bits of the sample as index */
  94         if (idx < (FIO_IO_U_PLAT_VAL << 1))
  95                 return idx;
  96
  97         /* Find the group and compute the minimum value of that group */
  98         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
  99         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 100
 101         /* Find its bucket number of the group */
 102         k = idx % FIO_IO_U_PLAT_VAL;
 103
 104         /* Return the mean of the range of the bucket */
 105         return base + ((k + 0.5) * (1 << error_bits));
 106 }
 107
 108 static int double_cmp(const void *a, const void *b)
 109 {
 110         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 111         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 112         int cmp = 0;
 113
 114         if (fa.u.f > fb.u.f)
 115                 cmp = 1;
 116         else if (fa.u.f < fb.u.f)
 117                 cmp = -1;
 118
 119         return cmp;
 120 }
 121
 122 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 123                                    fio_fp64_t *plist, unsigned int **output,
 124                                    unsigned int *maxv, unsigned int *minv)
 125 {
 126         unsigned long sum = 0;
 127         unsigned int len, i, j = 0;
 128         unsigned int oval_len = 0;
 129         unsigned int *ovals = NULL;
 130         int is_last;
 131
 132         *minv = -1U;
 133         *maxv = 0;
 134
 135         len = 0;
 136         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 137                 len++;
 138
 139         if (!len)
 140                 return 0;
 141
 142         /*
 143          * Sort the percentile list. Note that it may already be sorted if
 144          * we are using the default values, but since it's a short list this
 145          * isn't a worry. Also note that this does not work for NaN values.
 146          */
 147         if (len > 1)
 148                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 149
 150         /*
 151          * Calculate bucket values, note down max and min values
 152          */
 153         is_last = 0;
 154         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 155                 sum += io_u_plat[i];
 156                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 157                         assert(plist[j].u.f <= 100.0);
 158
 159                         if (j == oval_len) {
 160                                 oval_len += 100;
 161                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 162                         }
 163
 164                         ovals[j] = plat_idx_to_val(i);
 165                         if (ovals[j] < *minv)
 166                                 *minv = ovals[j];
 167                         if (ovals[j] > *maxv)
 168                                 *maxv = ovals[j];
 169
 170                         is_last = (j == len - 1);
 171                         if (is_last)
 172                                 break;
 173
 174                         j++;
 175                 }
 176         }
 177
 178         *output = ovals;
 179         return len;
 180 }
 181
 182 /*
 183  * Find and display the p-th percentile of clat
 184  */
 185 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 186                                   fio_fp64_t *plist, unsigned int precision)
 187 {
 188         unsigned int len, j = 0, minv, maxv;
 189         unsigned int *ovals;
 190         int is_last, per_line, scale_down;
 191         char fmt[32];
 192
 193         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 194         if (!len)
 195                 goto out;
 196
 197         /*
 198          * We default to usecs, but if the value range is such that we
 199          * should scale down to msecs, do that.
 200          */
 201         if (minv > 2000 && maxv > 99999) {
 202                 scale_down = 1;
 203                 log_info("    clat percentiles (msec):\n     |");
 204         } else {
 205                 scale_down = 0;
 206                 log_info("    clat percentiles (usec):\n     |");
 207         }
 208
 209         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 210         per_line = (80 - 7) / (precision + 14);
 211
 212         for (j = 0; j < len; j++) {
 213                 char fbuf[16], *ptr = fbuf;
 214
 215                 /* for formatting */
 216                 if (j != 0 && (j % per_line) == 0)
 217                         log_info("     |");
 218
 219                 /* end of the list */
 220                 is_last = (j == len - 1);
 221
 222                 if (plist[j].u.f < 10.0)
 223                         ptr += sprintf(fbuf, " ");
 224
 225                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 226
 227                 if (scale_down)
 228                         ovals[j] = (ovals[j] + 999) / 1000;
 229
 230                 log_info(" %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 231
 232                 if (is_last)
 233                         break;
 234
 235                 if ((j % per_line) == per_line - 1)     /* for formatting */
 236                         log_info("\n");
 237         }
 238
 239 out:
 240         if (ovals)
 241                 free(ovals);
 242 }
 243
 244 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 245              double *mean, double *dev)
 246 {
 247         double n = (double) is->samples;
 248
 249         if (n == 0)
 250                 return 0;
 251
 252         *min = is->min_val;
 253         *max = is->max_val;
 254         *mean = is->mean.u.f;
 255
 256         if (n > 1.0)
 257                 *dev = sqrt(is->S.u.f / (n - 1.0));
 258         else
 259                 *dev = 0;
 260
 261         return 1;
 262 }
 263
 264 void show_group_stats(struct group_run_stats *rs)
 265 {
 266         char *p1, *p2, *p3, *p4;
 267         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 268         int i;
 269
 270         log_info("\nRun status group %d (all jobs):\n", rs->groupid);
 271
 272         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 273                 const int i2p = is_power_of_2(rs->kb_base);
 274
 275                 if (!rs->max_run[i])
 276                         continue;
 277
 278                 p1 = num2str(rs->io_kb[i], 6, rs->kb_base, i2p, 8);
 279                 p2 = num2str(rs->agg[i], 6, rs->kb_base, i2p, rs->unit_base);
 280                 p3 = num2str(rs->min_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 281                 p4 = num2str(rs->max_bw[i], 6, rs->kb_base, i2p, rs->unit_base);
 282
 283                 log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
 284                          " mint=%llumsec, maxt=%llumsec\n",
 285                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 286                                 p1, p2, p3, p4,
 287                                 (unsigned long long) rs->min_run[i],
 288                                 (unsigned long long) rs->max_run[i]);
 289
 290                 free(p1);
 291                 free(p2);
 292                 free(p3);
 293                 free(p4);
 294         }
 295 }
 296
 297 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 298 {
 299         int i;
 300
 301         /*
 302          * Do depth distribution calculations
 303          */
 304         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 305                 if (total) {
 306                         io_u_dist[i] = (double) map[i] / (double) total;
 307                         io_u_dist[i] *= 100.0;
 308                         if (io_u_dist[i] < 0.1 && map[i])
 309                                 io_u_dist[i] = 0.1;
 310                 } else
 311                         io_u_dist[i] = 0.0;
 312         }
 313 }
 314
 315 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 316                           unsigned int *src, int nr)
 317 {
 318         unsigned long total = ddir_rw_sum(ts->total_io_u);
 319         int i;
 320
 321         /*
 322          * Do latency distribution calculations
 323          */
 324         for (i = 0; i < nr; i++) {
 325                 if (total) {
 326                         dst[i] = (double) src[i] / (double) total;
 327                         dst[i] *= 100.0;
 328                         if (dst[i] < 0.01 && src[i])
 329                                 dst[i] = 0.01;
 330                 } else
 331                         dst[i] = 0.0;
 332         }
 333 }
 334
 335 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 336 {
 337         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 338 }
 339
 340 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 341 {
 342         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 343 }
 344
 345 static void display_lat(const char *name, unsigned long min, unsigned long max,
 346                         double mean, double dev)
 347 {
 348         const char *base = "(usec)";
 349         char *minp, *maxp;
 350
 351         if (!usec_to_msec(&min, &max, &mean, &dev))
 352                 base = "(msec)";
 353
 354         minp = num2str(min, 6, 1, 0, 0);
 355         maxp = num2str(max, 6, 1, 0, 0);
 356
 357         log_info("    %s %s: min=%s, max=%s, avg=%5.02f,"
 358                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 359
 360         free(minp);
 361         free(maxp);
 362 }
 363
 364 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 365                              int ddir)
 366 {
 367         const char *str[] = { "read ", "write", "trim" };
 368         unsigned long min, max, runt;
 369         unsigned long long bw, iops;
 370         double mean, dev;
 371         char *io_p, *bw_p, *iops_p;
 372         int i2p;
 373
 374         assert(ddir_rw(ddir));
 375
 376         if (!ts->runtime[ddir])
 377                 return;
 378
 379         i2p = is_power_of_2(rs->kb_base);
 380         runt = ts->runtime[ddir];
 381
 382         bw = (1000 * ts->io_bytes[ddir]) / runt;
 383         io_p = num2str(ts->io_bytes[ddir], 6, 1, i2p, 8);
 384         bw_p = num2str(bw, 6, 1, i2p, ts->unit_base);
 385
 386         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 387         iops_p = num2str(iops, 6, 1, 0, 0);
 388
 389         log_info("  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
 390                                 rs->unified_rw_rep ? "mixed" : str[ddir],
 391                                 io_p, bw_p, iops_p,
 392                                 (unsigned long long) ts->runtime[ddir]);
 393
 394         free(io_p);
 395         free(bw_p);
 396         free(iops_p);
 397
 398         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 399                 display_lat("slat", min, max, mean, dev);
 400         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 401                 display_lat("clat", min, max, mean, dev);
 402         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 403                 display_lat(" lat", min, max, mean, dev);
 404
 405         if (ts->clat_percentiles) {
 406                 show_clat_percentiles(ts->io_u_plat[ddir],
 407                                         ts->clat_stat[ddir].samples,
 408                                         ts->percentile_list,
 409                                         ts->percentile_precision);
 410         }
 411         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 412                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 413                 const char *bw_str = (rs->unit_base == 1 ? "Kbit" : "KB");
 414
 415                 if (rs->unit_base == 1) {
 416                         min *= 8.0;
 417                         max *= 8.0;
 418                         mean *= 8.0;
 419                         dev *= 8.0;
 420                 }
 421
 422                 if (rs->agg[ddir]) {
 423                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 424                         if (p_of_agg > 100.0)
 425                                 p_of_agg = 100.0;
 426                 }
 427
 428                 if (mean > fkb_base * fkb_base) {
 429                         min /= fkb_base;
 430                         max /= fkb_base;
 431                         mean /= fkb_base;
 432                         dev /= fkb_base;
 433                         bw_str = (rs->unit_base == 1 ? "Mbit" : "MB");
 434                 }
 435
 436                 log_info("    bw (%-4s/s): min=%5lu, max=%5lu, per=%3.2f%%,"
 437                          " avg=%5.02f, stdev=%5.02f\n", bw_str, min, max,
 438                                                         p_of_agg, mean, dev);
 439         }
 440 }
 441
 442 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 443                     const char *msg)
 444 {
 445         int new_line = 1, i, line = 0, shown = 0;
 446
 447         for (i = 0; i < nr; i++) {
 448                 if (io_u_lat[i] <= 0.0)
 449                         continue;
 450                 shown = 1;
 451                 if (new_line) {
 452                         if (line)
 453                                 log_info("\n");
 454                         log_info("    lat (%s) : ", msg);
 455                         new_line = 0;
 456                         line = 0;
 457                 }
 458                 if (line)
 459                         log_info(", ");
 460                 log_info("%s%3.2f%%", ranges[i], io_u_lat[i]);
 461                 line++;
 462                 if (line == 5)
 463                         new_line = 1;
 464         }
 465
 466         if (shown)
 467                 log_info("\n");
 468
 469         return shown;
 470 }
 471
 472 static void show_lat_u(double *io_u_lat_u)
 473 {
 474         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 475                                  "250=", "500=", "750=", "1000=", };
 476
 477         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec");
 478 }
 479
 480 static void show_lat_m(double *io_u_lat_m)
 481 {
 482         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 483                                  "250=", "500=", "750=", "1000=", "2000=",
 484                                  ">=2000=", };
 485
 486         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec");
 487 }
 488
 489 static void show_latencies(struct thread_stat *ts)
 490 {
 491         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 492         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 493
 494         stat_calc_lat_u(ts, io_u_lat_u);
 495         stat_calc_lat_m(ts, io_u_lat_m);
 496
 497         show_lat_u(io_u_lat_u);
 498         show_lat_m(io_u_lat_m);
 499 }
 500
 501 static int block_state_category(int block_state)
 502 {
 503         switch (block_state) {
 504         case BLOCK_STATE_UNINIT:
 505                 return 0;
 506         case BLOCK_STATE_TRIMMED:
 507         case BLOCK_STATE_WRITTEN:
 508                 return 1;
 509         case BLOCK_STATE_WRITE_FAILURE:
 510         case BLOCK_STATE_TRIM_FAILURE:
 511                 return 2;
 512         default:
 513                 assert(0);
 514                 return -1;
 515         }
 516 }
 517
 518 static int compare_block_infos(const void *bs1, const void *bs2)
 519 {
 520         uint32_t block1 = *(uint32_t *)bs1;
 521         uint32_t block2 = *(uint32_t *)bs2;
 522         int state1 = BLOCK_INFO_STATE(block1);
 523         int state2 = BLOCK_INFO_STATE(block2);
 524         int bscat1 = block_state_category(state1);
 525         int bscat2 = block_state_category(state2);
 526         int cycles1 = BLOCK_INFO_TRIMS(block1);
 527         int cycles2 = BLOCK_INFO_TRIMS(block2);
 528
 529         if (bscat1 < bscat2)
 530                 return -1;
 531         if (bscat1 > bscat2)
 532                 return 1;
 533
 534         if (cycles1 < cycles2)
 535                 return -1;
 536         if (cycles1 > cycles2)
 537                 return 1;
 538
 539         if (state1 < state2)
 540                 return -1;
 541         if (state1 > state2)
 542                 return 1;
 543
 544         assert(block1 == block2);
 545         return 0;
 546 }
 547
 548 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 549                                   fio_fp64_t *plist, unsigned int **percentiles,
 550                                   unsigned int *types)
 551 {
 552         int len = 0;
 553         int i, nr_uninit;
 554
 555         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 556
 557         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 558                 len++;
 559
 560         if (!len)
 561                 return 0;
 562
 563         /*
 564          * Sort the percentile list. Note that it may already be sorted if
 565          * we are using the default values, but since it's a short list this
 566          * isn't a worry. Also note that this does not work for NaN values.
 567          */
 568         if (len > 1)
 569                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 570
 571         nr_uninit = 0;
 572         /* Start only after the uninit entries end */
 573         for (nr_uninit = 0;
 574              nr_uninit < nr_block_infos
 575                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 576              nr_uninit ++)
 577                 ;
 578
 579         if (nr_uninit == nr_block_infos)
 580                 return 0;
 581
 582         *percentiles = calloc(len, sizeof(**percentiles));
 583
 584         for (i = 0; i < len; i++) {
 585                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 586                                 + nr_uninit;
 587                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 588         }
 589
 590         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 591         for (i = 0; i < nr_block_infos; i++)
 592                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 593
 594         return len;
 595 }
 596
 597 static const char *block_state_names[] = {
 598         [BLOCK_STATE_UNINIT] = "unwritten",
 599         [BLOCK_STATE_TRIMMED] = "trimmed",
 600         [BLOCK_STATE_WRITTEN] = "written",
 601         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 602         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 603 };
 604
 605 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 606                              fio_fp64_t *plist)
 607 {
 608         int len, pos, i;
 609         unsigned int *percentiles = NULL;
 610         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 611
 612         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 613                                      &percentiles, block_state_counts);
 614
 615         log_info("  block lifetime percentiles :\n   |");
 616         pos = 0;
 617         for (i = 0; i < len; i++) {
 618                 uint32_t block_info = percentiles[i];
 619 #define LINE_LENGTH     75
 620                 char str[LINE_LENGTH];
 621                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 622                                      plist[i].u.f, block_info,
 623                                      i == len - 1 ? '\n' : ',');
 624                 assert(strln < LINE_LENGTH);
 625                 if (pos + strln > LINE_LENGTH) {
 626                         pos = 0;
 627                         log_info("\n   |");
 628                 }
 629                 log_info("%s", str);
 630                 pos += strln;
 631 #undef LINE_LENGTH
 632         }
 633         if (percentiles)
 634                 free(percentiles);
 635
 636         log_info("        states               :");
 637         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 638                 log_info(" %s=%u%c",
 639                          block_state_names[i], block_state_counts[i],
 640                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 641 }
 642
 643 static void show_thread_status_normal(struct thread_stat *ts,
 644                                       struct group_run_stats *rs)
 645 {
 646         double usr_cpu, sys_cpu;
 647         unsigned long runtime;
 648         double io_u_dist[FIO_IO_U_MAP_NR];
 649         time_t time_p;
 650         char time_buf[32];
 651
 652         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 653                 return;
 654
 655         time(&time_p);
 656         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 657
 658         if (!ts->error) {
 659                 log_info("%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 660                                         ts->name, ts->groupid, ts->members,
 661                                         ts->error, (int) ts->pid, time_buf);
 662         } else {
 663                 log_info("%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 664                                         ts->name, ts->groupid, ts->members,
 665                                         ts->error, ts->verror, (int) ts->pid,
 666                                         time_buf);
 667         }
 668
 669         if (strlen(ts->description))
 670                 log_info("  Description  : [%s]\n", ts->description);
 671
 672         if (ts->io_bytes[DDIR_READ])
 673                 show_ddir_status(rs, ts, DDIR_READ);
 674         if (ts->io_bytes[DDIR_WRITE])
 675                 show_ddir_status(rs, ts, DDIR_WRITE);
 676         if (ts->io_bytes[DDIR_TRIM])
 677                 show_ddir_status(rs, ts, DDIR_TRIM);
 678
 679         show_latencies(ts);
 680
 681         runtime = ts->total_run_time;
 682         if (runtime) {
 683                 double runt = (double) runtime;
 684
 685                 usr_cpu = (double) ts->usr_time * 100 / runt;
 686                 sys_cpu = (double) ts->sys_time * 100 / runt;
 687         } else {
 688                 usr_cpu = 0;
 689                 sys_cpu = 0;
 690         }
 691
 692         log_info("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 693                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 694                         (unsigned long long) ts->ctx,
 695                         (unsigned long long) ts->majf,
 696                         (unsigned long long) ts->minf);
 697
 698         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 699         log_info("  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 700                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 701                                         io_u_dist[1], io_u_dist[2],
 702                                         io_u_dist[3], io_u_dist[4],
 703                                         io_u_dist[5], io_u_dist[6]);
 704
 705         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 706         log_info("     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 707                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 708                                         io_u_dist[1], io_u_dist[2],
 709                                         io_u_dist[3], io_u_dist[4],
 710                                         io_u_dist[5], io_u_dist[6]);
 711         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 712         log_info("     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 713                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 714                                         io_u_dist[1], io_u_dist[2],
 715                                         io_u_dist[3], io_u_dist[4],
 716                                         io_u_dist[5], io_u_dist[6]);
 717         log_info("     issued    : total=r=%llu/w=%llu/d=%llu,"
 718                                  " short=r=%llu/w=%llu/d=%llu,"
 719                                  " drop=r=%llu/w=%llu/d=%llu\n",
 720                                         (unsigned long long) ts->total_io_u[0],
 721                                         (unsigned long long) ts->total_io_u[1],
 722                                         (unsigned long long) ts->total_io_u[2],
 723                                         (unsigned long long) ts->short_io_u[0],
 724                                         (unsigned long long) ts->short_io_u[1],
 725                                         (unsigned long long) ts->short_io_u[2],
 726                                         (unsigned long long) ts->drop_io_u[0],
 727                                         (unsigned long long) ts->drop_io_u[1],
 728                                         (unsigned long long) ts->drop_io_u[2]);
 729         if (ts->continue_on_error) {
 730                 log_info("     errors    : total=%llu, first_error=%d/<%s>\n",
 731                                         (unsigned long long)ts->total_err_count,
 732                                         ts->first_error,
 733                                         strerror(ts->first_error));
 734         }
 735         if (ts->latency_depth) {
 736                 log_info("     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 737                                         (unsigned long long)ts->latency_target,
 738                                         (unsigned long long)ts->latency_window,
 739                                         ts->latency_percentile.u.f,
 740                                         ts->latency_depth);
 741         }
 742
 743         if (ts->nr_block_infos)
 744                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 745                                   ts->percentile_list);
 746 }
 747
 748 static void show_ddir_status_terse(struct thread_stat *ts,
 749                                    struct group_run_stats *rs, int ddir)
 750 {
 751         unsigned long min, max;
 752         unsigned long long bw, iops;
 753         unsigned int *ovals = NULL;
 754         double mean, dev;
 755         unsigned int len, minv, maxv;
 756         int i;
 757
 758         assert(ddir_rw(ddir));
 759
 760         iops = bw = 0;
 761         if (ts->runtime[ddir]) {
 762                 uint64_t runt = ts->runtime[ddir];
 763
 764                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 765                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 766         }
 767
 768         log_info(";%llu;%llu;%llu;%llu",
 769                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 770                                         (unsigned long long) ts->runtime[ddir]);
 771
 772         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 773                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 774         else
 775                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 776
 777         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 778                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 779         else
 780                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 781
 782         if (ts->clat_percentiles) {
 783                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 784                                         ts->clat_stat[ddir].samples,
 785                                         ts->percentile_list, &ovals, &maxv,
 786                                         &minv);
 787         } else
 788                 len = 0;
 789
 790         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 791                 if (i >= len) {
 792                         log_info(";0%%=0");
 793                         continue;
 794                 }
 795                 log_info(";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 796         }
 797
 798         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 799                 log_info(";%lu;%lu;%f;%f", min, max, mean, dev);
 800         else
 801                 log_info(";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 802
 803         if (ovals)
 804                 free(ovals);
 805
 806         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 807                 double p_of_agg = 100.0;
 808
 809                 if (rs->agg[ddir]) {
 810                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 811                         if (p_of_agg > 100.0)
 812                                 p_of_agg = 100.0;
 813                 }
 814
 815                 log_info(";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 816         } else
 817                 log_info(";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 818 }
 819
 820 static void add_ddir_status_json(struct thread_stat *ts,
 821                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 822 {
 823         unsigned long min, max;
 824         unsigned long long bw;
 825         unsigned int *ovals = NULL;
 826         double mean, dev, iops;
 827         unsigned int len, minv, maxv;
 828         int i;
 829         const char *ddirname[] = {"read", "write", "trim"};
 830         struct json_object *dir_object, *tmp_object, *percentile_object;
 831         char buf[120];
 832         double p_of_agg = 100.0;
 833
 834         assert(ddir_rw(ddir));
 835
 836         if (ts->unified_rw_rep && ddir != DDIR_READ)
 837                 return;
 838
 839         dir_object = json_create_object();
 840         json_object_add_value_object(parent,
 841                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 842
 843         bw = 0;
 844         iops = 0.0;
 845         if (ts->runtime[ddir]) {
 846                 uint64_t runt = ts->runtime[ddir];
 847
 848                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
 849                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 850         }
 851
 852         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 853         json_object_add_value_int(dir_object, "bw", bw);
 854         json_object_add_value_float(dir_object, "iops", iops);
 855         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 856         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 857         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 858         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 859
 860         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 861                 min = max = 0;
 862                 mean = dev = 0.0;
 863         }
 864         tmp_object = json_create_object();
 865         json_object_add_value_object(dir_object, "slat", tmp_object);
 866         json_object_add_value_int(tmp_object, "min", min);
 867         json_object_add_value_int(tmp_object, "max", max);
 868         json_object_add_value_float(tmp_object, "mean", mean);
 869         json_object_add_value_float(tmp_object, "stddev", dev);
 870
 871         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 872                 min = max = 0;
 873                 mean = dev = 0.0;
 874         }
 875         tmp_object = json_create_object();
 876         json_object_add_value_object(dir_object, "clat", tmp_object);
 877         json_object_add_value_int(tmp_object, "min", min);
 878         json_object_add_value_int(tmp_object, "max", max);
 879         json_object_add_value_float(tmp_object, "mean", mean);
 880         json_object_add_value_float(tmp_object, "stddev", dev);
 881
 882         if (ts->clat_percentiles) {
 883                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 884                                         ts->clat_stat[ddir].samples,
 885                                         ts->percentile_list, &ovals, &maxv,
 886                                         &minv);
 887         } else
 888                 len = 0;
 889
 890         percentile_object = json_create_object();
 891         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 892         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 893                 if (i >= len) {
 894                         json_object_add_value_int(percentile_object, "0.00", 0);
 895                         continue;
 896                 }
 897                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 898                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 899         }
 900
 901         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 902                 min = max = 0;
 903                 mean = dev = 0.0;
 904         }
 905         tmp_object = json_create_object();
 906         json_object_add_value_object(dir_object, "lat", tmp_object);
 907         json_object_add_value_int(tmp_object, "min", min);
 908         json_object_add_value_int(tmp_object, "max", max);
 909         json_object_add_value_float(tmp_object, "mean", mean);
 910         json_object_add_value_float(tmp_object, "stddev", dev);
 911         if (ovals)
 912                 free(ovals);
 913
 914         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 915                 if (rs->agg[ddir]) {
 916                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 917                         if (p_of_agg > 100.0)
 918                                 p_of_agg = 100.0;
 919                 }
 920         } else {
 921                 min = max = 0;
 922                 p_of_agg = mean = dev = 0.0;
 923         }
 924         json_object_add_value_int(dir_object, "bw_min", min);
 925         json_object_add_value_int(dir_object, "bw_max", max);
 926         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
 927         json_object_add_value_float(dir_object, "bw_mean", mean);
 928         json_object_add_value_float(dir_object, "bw_dev", dev);
 929 }
 930
 931 static void show_thread_status_terse_v2(struct thread_stat *ts,
 932                                         struct group_run_stats *rs)
 933 {
 934         double io_u_dist[FIO_IO_U_MAP_NR];
 935         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 936         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 937         double usr_cpu, sys_cpu;
 938         int i;
 939
 940         /* General Info */
 941         log_info("2;%s;%d;%d", ts->name, ts->groupid, ts->error);
 942         /* Log Read Status */
 943         show_ddir_status_terse(ts, rs, DDIR_READ);
 944         /* Log Write Status */
 945         show_ddir_status_terse(ts, rs, DDIR_WRITE);
 946         /* Log Trim Status */
 947         show_ddir_status_terse(ts, rs, DDIR_TRIM);
 948
 949         /* CPU Usage */
 950         if (ts->total_run_time) {
 951                 double runt = (double) ts->total_run_time;
 952
 953                 usr_cpu = (double) ts->usr_time * 100 / runt;
 954                 sys_cpu = (double) ts->sys_time * 100 / runt;
 955         } else {
 956                 usr_cpu = 0;
 957                 sys_cpu = 0;
 958         }
 959
 960         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
 961                                                 (unsigned long long) ts->ctx,
 962                                                 (unsigned long long) ts->majf,
 963                                                 (unsigned long long) ts->minf);
 964
 965         /* Calc % distribution of IO depths, usecond, msecond latency */
 966         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 967         stat_calc_lat_u(ts, io_u_lat_u);
 968         stat_calc_lat_m(ts, io_u_lat_m);
 969
 970         /* Only show fixed 7 I/O depth levels*/
 971         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
 972                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
 973                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
 974
 975         /* Microsecond latency */
 976         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
 977                 log_info(";%3.2f%%", io_u_lat_u[i]);
 978         /* Millisecond latency */
 979         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
 980                 log_info(";%3.2f%%", io_u_lat_m[i]);
 981         /* Additional output if continue_on_error set - default off*/
 982         if (ts->continue_on_error)
 983                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
 984         log_info("\n");
 985
 986         /* Additional output if description is set */
 987         if (strlen(ts->description))
 988                 log_info(";%s", ts->description);
 989
 990         log_info("\n");
 991 }
 992
 993 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
 994                                            struct group_run_stats *rs, int ver)
 995 {
 996         double io_u_dist[FIO_IO_U_MAP_NR];
 997         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 998         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 999         double usr_cpu, sys_cpu;
1000         int i;
1001
1002         /* General Info */
1003         log_info("%d;%s;%s;%d;%d", ver, fio_version_string,
1004                                         ts->name, ts->groupid, ts->error);
1005         /* Log Read Status */
1006         show_ddir_status_terse(ts, rs, DDIR_READ);
1007         /* Log Write Status */
1008         show_ddir_status_terse(ts, rs, DDIR_WRITE);
1009         /* Log Trim Status */
1010         if (ver == 4)
1011                 show_ddir_status_terse(ts, rs, DDIR_TRIM);
1012
1013         /* CPU Usage */
1014         if (ts->total_run_time) {
1015                 double runt = (double) ts->total_run_time;
1016
1017                 usr_cpu = (double) ts->usr_time * 100 / runt;
1018                 sys_cpu = (double) ts->sys_time * 100 / runt;
1019         } else {
1020                 usr_cpu = 0;
1021                 sys_cpu = 0;
1022         }
1023
1024         log_info(";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1025                                                 (unsigned long long) ts->ctx,
1026                                                 (unsigned long long) ts->majf,
1027                                                 (unsigned long long) ts->minf);
1028
1029         /* Calc % distribution of IO depths, usecond, msecond latency */
1030         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1031         stat_calc_lat_u(ts, io_u_lat_u);
1032         stat_calc_lat_m(ts, io_u_lat_m);
1033
1034         /* Only show fixed 7 I/O depth levels*/
1035         log_info(";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1036                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1037                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1038
1039         /* Microsecond latency */
1040         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1041                 log_info(";%3.2f%%", io_u_lat_u[i]);
1042         /* Millisecond latency */
1043         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1044                 log_info(";%3.2f%%", io_u_lat_m[i]);
1045
1046         /* disk util stats, if any */
1047         show_disk_util(1, NULL);
1048
1049         /* Additional output if continue_on_error set - default off*/
1050         if (ts->continue_on_error)
1051                 log_info(";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1052
1053         /* Additional output if description is set */
1054         if (strlen(ts->description))
1055                 log_info(";%s", ts->description);
1056
1057         log_info("\n");
1058 }
1059
1060 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1061                                     struct group_run_stats *rs)
1062 {
1063         struct json_object *root, *tmp;
1064         double io_u_dist[FIO_IO_U_MAP_NR];
1065         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1066         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1067         double usr_cpu, sys_cpu;
1068         int i;
1069
1070         root = json_create_object();
1071         json_object_add_value_string(root, "jobname", ts->name);
1072         json_object_add_value_int(root, "groupid", ts->groupid);
1073         json_object_add_value_int(root, "error", ts->error);
1074
1075         add_ddir_status_json(ts, rs, DDIR_READ, root);
1076         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1077         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1078
1079         /* CPU Usage */
1080         if (ts->total_run_time) {
1081                 double runt = (double) ts->total_run_time;
1082
1083                 usr_cpu = (double) ts->usr_time * 100 / runt;
1084                 sys_cpu = (double) ts->sys_time * 100 / runt;
1085         } else {
1086                 usr_cpu = 0;
1087                 sys_cpu = 0;
1088         }
1089         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1090         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1091         json_object_add_value_int(root, "ctx", ts->ctx);
1092         json_object_add_value_int(root, "majf", ts->majf);
1093         json_object_add_value_int(root, "minf", ts->minf);
1094
1095
1096         /* Calc % distribution of IO depths, usecond, msecond latency */
1097         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1098         stat_calc_lat_u(ts, io_u_lat_u);
1099         stat_calc_lat_m(ts, io_u_lat_m);
1100
1101         tmp = json_create_object();
1102         json_object_add_value_object(root, "iodepth_level", tmp);
1103         /* Only show fixed 7 I/O depth levels*/
1104         for (i = 0; i < 7; i++) {
1105                 char name[20];
1106                 if (i < 6)
1107                         snprintf(name, 20, "%d", 1 << i);
1108                 else
1109                         snprintf(name, 20, ">=%d", 1 << i);
1110                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1111         }
1112
1113         tmp = json_create_object();
1114         json_object_add_value_object(root, "latency_us", tmp);
1115         /* Microsecond latency */
1116         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1117                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1118                                  "250", "500", "750", "1000", };
1119                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1120         }
1121         /* Millisecond latency */
1122         tmp = json_create_object();
1123         json_object_add_value_object(root, "latency_ms", tmp);
1124         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1125                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1126                                  "250", "500", "750", "1000", "2000",
1127                                  ">=2000", };
1128                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1129         }
1130
1131         /* Additional output if continue_on_error set - default off*/
1132         if (ts->continue_on_error) {
1133                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1134                 json_object_add_value_int(root, "first_error", ts->first_error);
1135         }
1136
1137         if (ts->latency_depth) {
1138                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1139                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1140                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1141                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1142         }
1143
1144         /* Additional output if description is set */
1145         if (strlen(ts->description))
1146                 json_object_add_value_string(root, "desc", ts->description);
1147
1148         if (ts->nr_block_infos) {
1149                 /* Block error histogram and types */
1150                 int len;
1151                 unsigned int *percentiles = NULL;
1152                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1153
1154                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1155                                              ts->percentile_list,
1156                                              &percentiles, block_state_counts);
1157
1158                 if (len) {
1159                         struct json_object *block, *percentile_object, *states;
1160                         int state, i;
1161                         block = json_create_object();
1162                         json_object_add_value_object(root, "block", block);
1163
1164                         percentile_object = json_create_object();
1165                         json_object_add_value_object(block, "percentiles",
1166                                                      percentile_object);
1167                         for (i = 0; i < len; i++) {
1168                                 char buf[20];
1169                                 snprintf(buf, sizeof(buf), "%f",
1170                                          ts->percentile_list[i].u.f);
1171                                 json_object_add_value_int(percentile_object,
1172                                                           (const char *)buf,
1173                                                           percentiles[i]);
1174                         }
1175
1176                         states = json_create_object();
1177                         json_object_add_value_object(block, "states", states);
1178                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1179                                 json_object_add_value_int(states,
1180                                         block_state_names[state],
1181                                         block_state_counts[state]);
1182                         }
1183                         free(percentiles);
1184                 }
1185         }
1186
1187         return root;
1188 }
1189
1190 static void show_thread_status_terse(struct thread_stat *ts,
1191                                      struct group_run_stats *rs)
1192 {
1193         if (terse_version == 2)
1194                 show_thread_status_terse_v2(ts, rs);
1195         else if (terse_version == 3 || terse_version == 4)
1196                 show_thread_status_terse_v3_v4(ts, rs, terse_version);
1197         else
1198                 log_err("fio: bad terse version!? %d\n", terse_version);
1199 }
1200
1201 struct json_object *show_thread_status(struct thread_stat *ts,
1202                                        struct group_run_stats *rs)
1203 {
1204         if (output_format == FIO_OUTPUT_TERSE)
1205                 show_thread_status_terse(ts, rs);
1206         else if (output_format == FIO_OUTPUT_JSON)
1207                 return show_thread_status_json(ts, rs);
1208         else
1209                 show_thread_status_normal(ts, rs);
1210         return NULL;
1211 }
1212
1213 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
1214 {
1215         double mean, S;
1216
1217         if (src->samples == 0)
1218                 return;
1219
1220         dst->min_val = min(dst->min_val, src->min_val);
1221         dst->max_val = max(dst->max_val, src->max_val);
1222
1223         /*
1224          * Compute new mean and S after the merge
1225          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1226          *  #Parallel_algorithm>
1227          */
1228         if (nr == 1) {
1229                 mean = src->mean.u.f;
1230                 S = src->S.u.f;
1231         } else {
1232                 double delta = src->mean.u.f - dst->mean.u.f;
1233
1234                 mean = ((src->mean.u.f * src->samples) +
1235                         (dst->mean.u.f * dst->samples)) /
1236                         (dst->samples + src->samples);
1237
1238                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1239                         (dst->samples * src->samples) /
1240                         (dst->samples + src->samples);
1241         }
1242
1243         dst->samples += src->samples;
1244         dst->mean.u.f = mean;
1245         dst->S.u.f = S;
1246 }
1247
1248 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1249 {
1250         int i;
1251
1252         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1253                 if (dst->max_run[i] < src->max_run[i])
1254                         dst->max_run[i] = src->max_run[i];
1255                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1256                         dst->min_run[i] = src->min_run[i];
1257                 if (dst->max_bw[i] < src->max_bw[i])
1258                         dst->max_bw[i] = src->max_bw[i];
1259                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1260                         dst->min_bw[i] = src->min_bw[i];
1261
1262                 dst->io_kb[i] += src->io_kb[i];
1263                 dst->agg[i] += src->agg[i];
1264         }
1265
1266         if (!dst->kb_base)
1267                 dst->kb_base = src->kb_base;
1268         if (!dst->unit_base)
1269                 dst->unit_base = src->unit_base;
1270 }
1271
1272 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
1273 {
1274         int l, k;
1275
1276         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1277                 if (!dst->unified_rw_rep) {
1278                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
1279                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
1280                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
1281                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
1282
1283                         dst->io_bytes[l] += src->io_bytes[l];
1284
1285                         if (dst->runtime[l] < src->runtime[l])
1286                                 dst->runtime[l] = src->runtime[l];
1287                 } else {
1288                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
1289                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
1290                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
1291                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
1292
1293                         dst->io_bytes[0] += src->io_bytes[l];
1294
1295                         if (dst->runtime[0] < src->runtime[l])
1296                                 dst->runtime[0] = src->runtime[l];
1297                 }
1298         }
1299
1300         dst->usr_time += src->usr_time;
1301         dst->sys_time += src->sys_time;
1302         dst->ctx += src->ctx;
1303         dst->majf += src->majf;
1304         dst->minf += src->minf;
1305
1306         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1307                 dst->io_u_map[k] += src->io_u_map[k];
1308         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1309                 dst->io_u_submit[k] += src->io_u_submit[k];
1310         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1311                 dst->io_u_complete[k] += src->io_u_complete[k];
1312         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1313                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1314         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1315                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1316
1317         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1318                 if (!dst->unified_rw_rep) {
1319                         dst->total_io_u[k] += src->total_io_u[k];
1320                         dst->short_io_u[k] += src->short_io_u[k];
1321                         dst->drop_io_u[k] += src->drop_io_u[k];
1322                 } else {
1323                         dst->total_io_u[0] += src->total_io_u[k];
1324                         dst->short_io_u[0] += src->short_io_u[k];
1325                         dst->drop_io_u[0] += src->drop_io_u[k];
1326                 }
1327         }
1328
1329         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1330                 int m;
1331
1332                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1333                         if (!dst->unified_rw_rep)
1334                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1335                         else
1336                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1337                 }
1338         }
1339
1340         dst->total_run_time += src->total_run_time;
1341         dst->total_submit += src->total_submit;
1342         dst->total_complete += src->total_complete;
1343 }
1344
1345 void init_group_run_stat(struct group_run_stats *gs)
1346 {
1347         int i;
1348         memset(gs, 0, sizeof(*gs));
1349
1350         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1351                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1352 }
1353
1354 void init_thread_stat(struct thread_stat *ts)
1355 {
1356         int j;
1357
1358         memset(ts, 0, sizeof(*ts));
1359
1360         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1361                 ts->lat_stat[j].min_val = -1UL;
1362                 ts->clat_stat[j].min_val = -1UL;
1363                 ts->slat_stat[j].min_val = -1UL;
1364                 ts->bw_stat[j].min_val = -1UL;
1365         }
1366         ts->groupid = -1;
1367 }
1368
1369 void __show_run_stats(void)
1370 {
1371         struct group_run_stats *runstats, *rs;
1372         struct thread_data *td;
1373         struct thread_stat *threadstats, *ts;
1374         int i, j, k, nr_ts, last_ts, idx;
1375         int kb_base_warned = 0;
1376         int unit_base_warned = 0;
1377         struct json_object *root = NULL;
1378         struct json_array *array = NULL;
1379         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1380
1381         for (i = 0; i < groupid + 1; i++)
1382                 init_group_run_stat(&runstats[i]);
1383
1384         /*
1385          * find out how many threads stats we need. if group reporting isn't
1386          * enabled, it's one-per-td.
1387          */
1388         nr_ts = 0;
1389         last_ts = -1;
1390         for_each_td(td, i) {
1391                 if (!td->o.group_reporting) {
1392                         nr_ts++;
1393                         continue;
1394                 }
1395                 if (last_ts == td->groupid)
1396                         continue;
1397
1398                 last_ts = td->groupid;
1399                 nr_ts++;
1400         }
1401
1402         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1403
1404         for (i = 0; i < nr_ts; i++)
1405                 init_thread_stat(&threadstats[i]);
1406
1407         j = 0;
1408         last_ts = -1;
1409         idx = 0;
1410         for_each_td(td, i) {
1411                 if (idx && (!td->o.group_reporting ||
1412                     (td->o.group_reporting && last_ts != td->groupid))) {
1413                         idx = 0;
1414                         j++;
1415                 }
1416
1417                 last_ts = td->groupid;
1418
1419                 ts = &threadstats[j];
1420
1421                 ts->clat_percentiles = td->o.clat_percentiles;
1422                 ts->percentile_precision = td->o.percentile_precision;
1423                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1424
1425                 idx++;
1426                 ts->members++;
1427
1428                 if (ts->groupid == -1) {
1429                         /*
1430                          * These are per-group shared already
1431                          */
1432                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1433                         if (td->o.description)
1434                                 strncpy(ts->description, td->o.description,
1435                                                 FIO_JOBDESC_SIZE - 1);
1436                         else
1437                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1438
1439                         /*
1440                          * If multiple entries in this group, this is
1441                          * the first member.
1442                          */
1443                         ts->thread_number = td->thread_number;
1444                         ts->groupid = td->groupid;
1445
1446                         /*
1447                          * first pid in group, not very useful...
1448                          */
1449                         ts->pid = td->pid;
1450
1451                         ts->kb_base = td->o.kb_base;
1452                         ts->unit_base = td->o.unit_base;
1453                         ts->unified_rw_rep = td->o.unified_rw_rep;
1454                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1455                         log_info("fio: kb_base differs for jobs in group, using"
1456                                  " %u as the base\n", ts->kb_base);
1457                         kb_base_warned = 1;
1458                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1459                         log_info("fio: unit_base differs for jobs in group, using"
1460                                  " %u as the base\n", ts->unit_base);
1461                         unit_base_warned = 1;
1462                 }
1463
1464                 ts->continue_on_error = td->o.continue_on_error;
1465                 ts->total_err_count += td->total_err_count;
1466                 ts->first_error = td->first_error;
1467                 if (!ts->error) {
1468                         if (!td->error && td->o.continue_on_error &&
1469                             td->first_error) {
1470                                 ts->error = td->first_error;
1471                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1472                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1473                         } else  if (td->error) {
1474                                 ts->error = td->error;
1475                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1476                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1477                         }
1478                 }
1479
1480                 ts->latency_depth = td->latency_qd;
1481                 ts->latency_target = td->o.latency_target;
1482                 ts->latency_percentile = td->o.latency_percentile;
1483                 ts->latency_window = td->o.latency_window;
1484
1485                 ts->nr_block_infos = td->ts.nr_block_infos;
1486                 for (k = 0; k < ts->nr_block_infos; k++)
1487                         ts->block_infos[k] = td->ts.block_infos[k];
1488
1489                 sum_thread_stats(ts, &td->ts, idx);
1490         }
1491
1492         for (i = 0; i < nr_ts; i++) {
1493                 unsigned long long bw;
1494
1495                 ts = &threadstats[i];
1496                 rs = &runstats[ts->groupid];
1497                 rs->kb_base = ts->kb_base;
1498                 rs->unit_base = ts->unit_base;
1499                 rs->unified_rw_rep += ts->unified_rw_rep;
1500
1501                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1502                         if (!ts->runtime[j])
1503                                 continue;
1504                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1505                                 rs->min_run[j] = ts->runtime[j];
1506                         if (ts->runtime[j] > rs->max_run[j])
1507                                 rs->max_run[j] = ts->runtime[j];
1508
1509                         bw = 0;
1510                         if (ts->runtime[j]) {
1511                                 unsigned long runt = ts->runtime[j];
1512                                 unsigned long long kb;
1513
1514                                 kb = ts->io_bytes[j] / rs->kb_base;
1515                                 bw = kb * 1000 / runt;
1516                         }
1517                         if (bw < rs->min_bw[j])
1518                                 rs->min_bw[j] = bw;
1519                         if (bw > rs->max_bw[j])
1520                                 rs->max_bw[j] = bw;
1521
1522                         rs->io_kb[j] += ts->io_bytes[j] / rs->kb_base;
1523                 }
1524         }
1525
1526         for (i = 0; i < groupid + 1; i++) {
1527                 int ddir;
1528
1529                 rs = &runstats[i];
1530
1531                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1532                         if (rs->max_run[ddir])
1533                                 rs->agg[ddir] = (rs->io_kb[ddir] * 1000) /
1534                                                 rs->max_run[ddir];
1535                 }
1536         }
1537
1538         /*
1539          * don't overwrite last signal output
1540          */
1541         if (output_format == FIO_OUTPUT_NORMAL)
1542                 log_info("\n");
1543         else if (output_format == FIO_OUTPUT_JSON) {
1544                 char time_buf[32];
1545                 time_t time_p;
1546
1547                 time(&time_p);
1548                 os_ctime_r((const time_t *) &time_p, time_buf,
1549                                 sizeof(time_buf));
1550                 time_buf[strlen(time_buf) - 1] = '\0';
1551
1552                 root = json_create_object();
1553                 json_object_add_value_string(root, "fio version", fio_version_string);
1554                 json_object_add_value_int(root, "timestamp", time_p);
1555                 json_object_add_value_string(root, "time", time_buf);
1556                 array = json_create_array();
1557                 json_object_add_value_array(root, "jobs", array);
1558         }
1559
1560         for (i = 0; i < nr_ts; i++) {
1561                 ts = &threadstats[i];
1562                 rs = &runstats[ts->groupid];
1563
1564                 if (is_backend)
1565                         fio_server_send_ts(ts, rs);
1566                 else if (output_format == FIO_OUTPUT_TERSE)
1567                         show_thread_status_terse(ts, rs);
1568                 else if (output_format == FIO_OUTPUT_JSON) {
1569                         struct json_object *tmp = show_thread_status_json(ts, rs);
1570                         json_array_add_value_object(array, tmp);
1571                 } else
1572                         show_thread_status_normal(ts, rs);
1573         }
1574         if (output_format == FIO_OUTPUT_JSON) {
1575                 /* disk util stats, if any */
1576                 show_disk_util(1, root);
1577
1578                 show_idle_prof_stats(FIO_OUTPUT_JSON, root);
1579
1580                 json_print_object(root);
1581                 log_info("\n");
1582                 json_free_object(root);
1583         }
1584
1585         for (i = 0; i < groupid + 1; i++) {
1586                 rs = &runstats[i];
1587
1588                 rs->groupid = i;
1589                 if (is_backend)
1590                         fio_server_send_gs(rs);
1591                 else if (output_format == FIO_OUTPUT_NORMAL)
1592                         show_group_stats(rs);
1593         }
1594
1595         if (is_backend)
1596                 fio_server_send_du();
1597         else if (output_format == FIO_OUTPUT_NORMAL) {
1598                 show_disk_util(0, NULL);
1599                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
1600         }
1601
1602         if ( !(output_format == FIO_OUTPUT_TERSE) && append_terse_output) {
1603                 log_info("\nAdditional Terse Output:\n");
1604
1605                 for (i = 0; i < nr_ts; i++) {
1606                         ts = &threadstats[i];
1607                         rs = &runstats[ts->groupid];
1608                         show_thread_status_terse(ts, rs);
1609                 }
1610         }
1611
1612         log_info_flush();
1613         free(runstats);
1614         free(threadstats);
1615 }
1616
1617 void show_run_stats(void)
1618 {
1619         fio_mutex_down(stat_mutex);
1620         __show_run_stats();
1621         fio_mutex_up(stat_mutex);
1622 }
1623
1624 void __show_running_run_stats(void)
1625 {
1626         struct thread_data *td;
1627         unsigned long long *rt;
1628         struct timeval tv;
1629         int i;
1630
1631         fio_mutex_down(stat_mutex);
1632
1633         rt = malloc(thread_number * sizeof(unsigned long long));
1634         fio_gettime(&tv, NULL);
1635
1636         for_each_td(td, i) {
1637                 rt[i] = mtime_since(&td->start, &tv);
1638                 if (td_read(td) && td->io_bytes[DDIR_READ])
1639                         td->ts.runtime[DDIR_READ] += rt[i];
1640                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1641                         td->ts.runtime[DDIR_WRITE] += rt[i];
1642                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1643                         td->ts.runtime[DDIR_TRIM] += rt[i];
1644
1645                 td->update_rusage = 1;
1646                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1647                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1648                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1649                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1650         }
1651
1652         for_each_td(td, i) {
1653                 if (td->runstate >= TD_EXITED)
1654                         continue;
1655                 if (td->rusage_sem) {
1656                         td->update_rusage = 1;
1657                         fio_mutex_down(td->rusage_sem);
1658                 }
1659                 td->update_rusage = 0;
1660         }
1661
1662         __show_run_stats();
1663
1664         for_each_td(td, i) {
1665                 if (td_read(td) && td->io_bytes[DDIR_READ])
1666                         td->ts.runtime[DDIR_READ] -= rt[i];
1667                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1668                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1669                 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1670                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1671         }
1672
1673         free(rt);
1674         fio_mutex_up(stat_mutex);
1675 }
1676
1677 static int status_interval_init;
1678 static struct timeval status_time;
1679 static int status_file_disabled;
1680
1681 #define FIO_STATUS_FILE         "fio-dump-status"
1682
1683 static int check_status_file(void)
1684 {
1685         struct stat sb;
1686         const char *temp_dir;
1687         char fio_status_file_path[PATH_MAX];
1688
1689         if (status_file_disabled)
1690                 return 0;
1691
1692         temp_dir = getenv("TMPDIR");
1693         if (temp_dir == NULL) {
1694                 temp_dir = getenv("TEMP");
1695                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1696                         temp_dir = NULL;
1697         }
1698         if (temp_dir == NULL)
1699                 temp_dir = "/tmp";
1700
1701         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1702
1703         if (stat(fio_status_file_path, &sb))
1704                 return 0;
1705
1706         if (unlink(fio_status_file_path) < 0) {
1707                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1708                                                         strerror(errno));
1709                 log_err("fio: disabling status file updates\n");
1710                 status_file_disabled = 1;
1711         }
1712
1713         return 1;
1714 }
1715
1716 void check_for_running_stats(void)
1717 {
1718         if (status_interval) {
1719                 if (!status_interval_init) {
1720                         fio_gettime(&status_time, NULL);
1721                         status_interval_init = 1;
1722                 } else if (mtime_since_now(&status_time) >= status_interval) {
1723                         show_running_run_stats();
1724                         fio_gettime(&status_time, NULL);
1725                         return;
1726                 }
1727         }
1728         if (check_status_file()) {
1729                 show_running_run_stats();
1730                 return;
1731         }
1732 }
1733
1734 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1735 {
1736         double val = data;
1737         double delta;
1738
1739         if (data > is->max_val)
1740                 is->max_val = data;
1741         if (data < is->min_val)
1742                 is->min_val = data;
1743
1744         delta = val - is->mean.u.f;
1745         if (delta) {
1746                 is->mean.u.f += delta / (is->samples + 1.0);
1747                 is->S.u.f += delta * (val - is->mean.u.f);
1748         }
1749
1750         is->samples++;
1751 }
1752
1753 static void __add_log_sample(struct io_log *iolog, unsigned long val,
1754                              enum fio_ddir ddir, unsigned int bs,
1755                              unsigned long t, uint64_t offset)
1756 {
1757         uint64_t nr_samples = iolog->nr_samples;
1758         struct io_sample *s;
1759
1760         if (iolog->disabled)
1761                 return;
1762
1763         if (!iolog->nr_samples)
1764                 iolog->avg_last = t;
1765
1766         if (iolog->nr_samples == iolog->max_samples) {
1767                 size_t new_size;
1768                 void *new_log;
1769
1770                 new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
1771
1772                 if (iolog->log_gz && (new_size > iolog->log_gz)) {
1773                         if (iolog_flush(iolog, 0)) {
1774                                 log_err("fio: failed flushing iolog! Will stop logging.\n");
1775                                 iolog->disabled = 1;
1776                                 return;
1777                         }
1778                         nr_samples = iolog->nr_samples;
1779                 } else {
1780                         new_log = realloc(iolog->log, new_size);
1781                         if (!new_log) {
1782                                 log_err("fio: failed extending iolog! Will stop logging.\n");
1783                                 iolog->disabled = 1;
1784                                 return;
1785                         }
1786                         iolog->log = new_log;
1787                         iolog->max_samples <<= 1;
1788                 }
1789         }
1790
1791         s = get_sample(iolog, nr_samples);
1792
1793         s->val = val;
1794         s->time = t;
1795         io_sample_set_ddir(iolog, s, ddir);
1796         s->bs = bs;
1797
1798         if (iolog->log_offset) {
1799                 struct io_sample_offset *so = (void *) s;
1800
1801                 so->offset = offset;
1802         }
1803
1804         iolog->nr_samples++;
1805 }
1806
1807 static inline void reset_io_stat(struct io_stat *ios)
1808 {
1809         ios->max_val = ios->min_val = ios->samples = 0;
1810         ios->mean.u.f = ios->S.u.f = 0;
1811 }
1812
1813 void reset_io_stats(struct thread_data *td)
1814 {
1815         struct thread_stat *ts = &td->ts;
1816         int i, j;
1817
1818         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1819                 reset_io_stat(&ts->clat_stat[i]);
1820                 reset_io_stat(&ts->slat_stat[i]);
1821                 reset_io_stat(&ts->lat_stat[i]);
1822                 reset_io_stat(&ts->bw_stat[i]);
1823                 reset_io_stat(&ts->iops_stat[i]);
1824
1825                 ts->io_bytes[i] = 0;
1826                 ts->runtime[i] = 0;
1827
1828                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1829                         ts->io_u_plat[i][j] = 0;
1830         }
1831
1832         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1833                 ts->io_u_map[i] = 0;
1834                 ts->io_u_submit[i] = 0;
1835                 ts->io_u_complete[i] = 0;
1836                 ts->io_u_lat_u[i] = 0;
1837                 ts->io_u_lat_m[i] = 0;
1838                 ts->total_submit = 0;
1839                 ts->total_complete = 0;
1840         }
1841
1842         for (i = 0; i < 3; i++) {
1843                 ts->total_io_u[i] = 0;
1844                 ts->short_io_u[i] = 0;
1845                 ts->drop_io_u[i] = 0;
1846         }
1847 }
1848
1849 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
1850 {
1851         /*
1852          * Note an entry in the log. Use the mean from the logged samples,
1853          * making sure to properly round up. Only write a log entry if we
1854          * had actual samples done.
1855          */
1856         if (iolog->avg_window[DDIR_READ].samples) {
1857                 unsigned long mr;
1858
1859                 mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
1860                 __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
1861         }
1862         if (iolog->avg_window[DDIR_WRITE].samples) {
1863                 unsigned long mw;
1864
1865                 mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
1866                 __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
1867         }
1868         if (iolog->avg_window[DDIR_TRIM].samples) {
1869                 unsigned long mw;
1870
1871                 mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
1872                 __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
1873         }
1874
1875         reset_io_stat(&iolog->avg_window[DDIR_READ]);
1876         reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
1877         reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
1878 }
1879
1880 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
1881                            unsigned long val, enum fio_ddir ddir,
1882                            unsigned int bs, uint64_t offset)
1883 {
1884         unsigned long elapsed, this_window;
1885
1886         if (!ddir_rw(ddir))
1887                 return;
1888
1889         elapsed = mtime_since_now(&td->epoch);
1890
1891         /*
1892          * If no time averaging, just add the log sample.
1893          */
1894         if (!iolog->avg_msec) {
1895                 __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
1896                 return;
1897         }
1898
1899         /*
1900          * Add the sample. If the time period has passed, then
1901          * add that entry to the log and clear.
1902          */
1903         add_stat_sample(&iolog->avg_window[ddir], val);
1904
1905         /*
1906          * If period hasn't passed, adding the above sample is all we
1907          * need to do.
1908          */
1909         this_window = elapsed - iolog->avg_last;
1910         if (this_window < iolog->avg_msec)
1911                 return;
1912
1913         _add_stat_to_log(iolog, elapsed);
1914
1915         iolog->avg_last = elapsed;
1916 }
1917
1918 void finalize_logs(struct thread_data *td)
1919 {
1920         unsigned long elapsed;
1921
1922         elapsed = mtime_since_now(&td->epoch);
1923
1924         if (td->clat_log)
1925                 _add_stat_to_log(td->clat_log, elapsed);
1926         if (td->slat_log)
1927                 _add_stat_to_log(td->slat_log, elapsed);
1928         if (td->lat_log)
1929                 _add_stat_to_log(td->lat_log, elapsed);
1930         if (td->bw_log)
1931                 _add_stat_to_log(td->bw_log, elapsed);
1932         if (td->iops_log)
1933                 _add_stat_to_log(td->iops_log, elapsed);
1934 }
1935
1936 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
1937 {
1938         struct io_log *iolog;
1939
1940         if (!ddir_rw(ddir))
1941                 return;
1942
1943         iolog = agg_io_log[ddir];
1944         __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
1945 }
1946
1947 static void add_clat_percentile_sample(struct thread_stat *ts,
1948                                 unsigned long usec, enum fio_ddir ddir)
1949 {
1950         unsigned int idx = plat_val_to_idx(usec);
1951         assert(idx < FIO_IO_U_PLAT_NR);
1952
1953         ts->io_u_plat[ddir][idx]++;
1954 }
1955
1956 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
1957                      unsigned long usec, unsigned int bs, uint64_t offset)
1958 {
1959         struct thread_stat *ts = &td->ts;
1960
1961         if (!ddir_rw(ddir))
1962                 return;
1963
1964         add_stat_sample(&ts->clat_stat[ddir], usec);
1965
1966         if (td->clat_log)
1967                 add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
1968
1969         if (ts->clat_percentiles)
1970                 add_clat_percentile_sample(ts, usec, ddir);
1971 }
1972
1973 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
1974                      unsigned long usec, unsigned int bs, uint64_t offset)
1975 {
1976         struct thread_stat *ts = &td->ts;
1977
1978         if (!ddir_rw(ddir))
1979                 return;
1980
1981         add_stat_sample(&ts->slat_stat[ddir], usec);
1982
1983         if (td->slat_log)
1984                 add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
1985 }
1986
1987 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
1988                     unsigned long usec, unsigned int bs, uint64_t offset)
1989 {
1990         struct thread_stat *ts = &td->ts;
1991
1992         if (!ddir_rw(ddir))
1993                 return;
1994
1995         add_stat_sample(&ts->lat_stat[ddir], usec);
1996
1997         if (td->lat_log)
1998                 add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
1999 }
2000
2001 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2002                    struct timeval *t)
2003 {
2004         struct thread_stat *ts = &td->ts;
2005         unsigned long spent, rate;
2006
2007         if (!ddir_rw(ddir))
2008                 return;
2009
2010         spent = mtime_since(&td->bw_sample_time, t);
2011         if (spent < td->o.bw_avg_time)
2012                 return;
2013
2014         td_io_u_lock(td);
2015
2016         /*
2017          * Compute both read and write rates for the interval.
2018          */
2019         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2020                 uint64_t delta;
2021
2022                 delta = td->this_io_bytes[ddir] - td->stat_io_bytes[ddir];
2023                 if (!delta)
2024                         continue; /* No entries for interval */
2025
2026                 if (spent)
2027                         rate = delta * 1000 / spent / 1024;
2028                 else
2029                         rate = 0;
2030
2031                 add_stat_sample(&ts->bw_stat[ddir], rate);
2032
2033                 if (td->bw_log)
2034                         add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
2035
2036                 td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
2037         }
2038
2039         fio_gettime(&td->bw_sample_time, NULL);
2040         td_io_u_unlock(td);
2041 }
2042
2043 void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
2044                      struct timeval *t)
2045 {
2046         struct thread_stat *ts = &td->ts;
2047         unsigned long spent, iops;
2048
2049         if (!ddir_rw(ddir))
2050                 return;
2051
2052         spent = mtime_since(&td->iops_sample_time, t);
2053         if (spent < td->o.iops_avg_time)
2054                 return;
2055
2056         td_io_u_lock(td);
2057
2058         /*
2059          * Compute both read and write rates for the interval.
2060          */
2061         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
2062                 uint64_t delta;
2063
2064                 delta = td->this_io_blocks[ddir] - td->stat_io_blocks[ddir];
2065                 if (!delta)
2066                         continue; /* No entries for interval */
2067
2068                 if (spent)
2069                         iops = (delta * 1000) / spent;
2070                 else
2071                         iops = 0;
2072
2073                 add_stat_sample(&ts->iops_stat[ddir], iops);
2074
2075                 if (td->iops_log)
2076                         add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
2077
2078                 td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
2079         }
2080
2081         fio_gettime(&td->iops_sample_time, NULL);
2082         td_io_u_unlock(td);
2083 }
2084
2085 void stat_init(void)
2086 {
2087         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2088 }
2089
2090 void stat_exit(void)
2091 {
2092         /*
2093          * When we have the mutex, we know out-of-band access to it
2094          * have ended.
2095          */
2096         fio_mutex_down(stat_mutex);
2097         fio_mutex_remove(stat_mutex);
2098 }
2099
2100 /*
2101  * Called from signal handler. Wake up status thread.
2102  */
2103 void show_running_run_stats(void)
2104 {
2105         helper_do_stat = 1;
2106         pthread_cond_signal(&helper_cond);
2107 }
2108
2109 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2110 {
2111         /* Ignore io_u's which span multiple blocks--they will just get
2112          * inaccurate counts. */
2113         int idx = (io_u->offset - io_u->file->file_offset)
2114                         / td->o.bs[DDIR_TRIM];
2115         uint32_t *info = &td->ts.block_infos[idx];
2116         assert(idx < td->ts.nr_block_infos);
2117         return info;
2118 }