stat.c

   1 #include <stdio.h>
   2 #include <string.h>
   3 #include <sys/time.h>
   4 #include <sys/types.h>
   5 #include <sys/stat.h>
   6 #include <dirent.h>
   7 #include <libgen.h>
   8 #include <math.h>
   9
  10 #include "fio.h"
  11 #include "diskutil.h"
  12 #include "lib/ieee754.h"
  13 #include "json.h"
  14 #include "lib/getrusage.h"
  15 #include "idletime.h"
  16 #include "lib/pow2.h"
  17 #include "lib/output_buffer.h"
  18 #include "helper_thread.h"
  19 #include "smalloc.h"
  20
  21 #define LOG_MSEC_SLACK  10
  22
  23 struct fio_mutex *stat_mutex;
  24
  25 void clear_rusage_stat(struct thread_data *td)
  26 {
  27         struct thread_stat *ts = &td->ts;
  28
  29         fio_getrusage(&td->ru_start);
  30         ts->usr_time = ts->sys_time = 0;
  31         ts->ctx = 0;
  32         ts->minf = ts->majf = 0;
  33 }
  34
  35 void update_rusage_stat(struct thread_data *td)
  36 {
  37         struct thread_stat *ts = &td->ts;
  38
  39         fio_getrusage(&td->ru_end);
  40         ts->usr_time += mtime_since(&td->ru_start.ru_utime,
  41                                         &td->ru_end.ru_utime);
  42         ts->sys_time += mtime_since(&td->ru_start.ru_stime,
  43                                         &td->ru_end.ru_stime);
  44         ts->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw
  45                         - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
  46         ts->minf += td->ru_end.ru_minflt - td->ru_start.ru_minflt;
  47         ts->majf += td->ru_end.ru_majflt - td->ru_start.ru_majflt;
  48
  49         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
  50 }
  51
  52 /*
  53  * Given a latency, return the index of the corresponding bucket in
  54  * the structure tracking percentiles.
  55  *
  56  * (1) find the group (and error bits) that the value (latency)
  57  * belongs to by looking at its MSB. (2) find the bucket number in the
  58  * group by looking at the index bits.
  59  *
  60  */
  61 static unsigned int plat_val_to_idx(unsigned int val)
  62 {
  63         unsigned int msb, error_bits, base, offset, idx;
  64
  65         /* Find MSB starting from bit 0 */
  66         if (val == 0)
  67                 msb = 0;
  68         else
  69                 msb = (sizeof(val)*8) - __builtin_clz(val) - 1;
  70
  71         /*
  72          * MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
  73          * all bits of the sample as index
  74          */
  75         if (msb <= FIO_IO_U_PLAT_BITS)
  76                 return val;
  77
  78         /* Compute the number of error bits to discard*/
  79         error_bits = msb - FIO_IO_U_PLAT_BITS;
  80
  81         /* Compute the number of buckets before the group */
  82         base = (error_bits + 1) << FIO_IO_U_PLAT_BITS;
  83
  84         /*
  85          * Discard the error bits and apply the mask to find the
  86          * index for the buckets in the group
  87          */
  88         offset = (FIO_IO_U_PLAT_VAL - 1) & (val >> error_bits);
  89
  90         /* Make sure the index does not exceed (array size - 1) */
  91         idx = (base + offset) < (FIO_IO_U_PLAT_NR - 1) ?
  92                 (base + offset) : (FIO_IO_U_PLAT_NR - 1);
  93
  94         return idx;
  95 }
  96
  97 /*
  98  * Convert the given index of the bucket array to the value
  99  * represented by the bucket
 100  */
 101 static unsigned long long plat_idx_to_val(unsigned int idx)
 102 {
 103         unsigned int error_bits, k, base;
 104
 105         assert(idx < FIO_IO_U_PLAT_NR);
 106
 107         /* MSB <= (FIO_IO_U_PLAT_BITS-1), cannot be rounded off. Use
 108          * all bits of the sample as index */
 109         if (idx < (FIO_IO_U_PLAT_VAL << 1))
 110                 return idx;
 111
 112         /* Find the group and compute the minimum value of that group */
 113         error_bits = (idx >> FIO_IO_U_PLAT_BITS) - 1;
 114         base = 1 << (error_bits + FIO_IO_U_PLAT_BITS);
 115
 116         /* Find its bucket number of the group */
 117         k = idx % FIO_IO_U_PLAT_VAL;
 118
 119         /* Return the mean of the range of the bucket */
 120         return base + ((k + 0.5) * (1 << error_bits));
 121 }
 122
 123 static int double_cmp(const void *a, const void *b)
 124 {
 125         const fio_fp64_t fa = *(const fio_fp64_t *) a;
 126         const fio_fp64_t fb = *(const fio_fp64_t *) b;
 127         int cmp = 0;
 128
 129         if (fa.u.f > fb.u.f)
 130                 cmp = 1;
 131         else if (fa.u.f < fb.u.f)
 132                 cmp = -1;
 133
 134         return cmp;
 135 }
 136
 137 unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 138                                    fio_fp64_t *plist, unsigned int **output,
 139                                    unsigned int *maxv, unsigned int *minv)
 140 {
 141         unsigned long sum = 0;
 142         unsigned int len, i, j = 0;
 143         unsigned int oval_len = 0;
 144         unsigned int *ovals = NULL;
 145         int is_last;
 146
 147         *minv = -1U;
 148         *maxv = 0;
 149
 150         len = 0;
 151         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 152                 len++;
 153
 154         if (!len)
 155                 return 0;
 156
 157         /*
 158          * Sort the percentile list. Note that it may already be sorted if
 159          * we are using the default values, but since it's a short list this
 160          * isn't a worry. Also note that this does not work for NaN values.
 161          */
 162         if (len > 1)
 163                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 164
 165         /*
 166          * Calculate bucket values, note down max and min values
 167          */
 168         is_last = 0;
 169         for (i = 0; i < FIO_IO_U_PLAT_NR && !is_last; i++) {
 170                 sum += io_u_plat[i];
 171                 while (sum >= (plist[j].u.f / 100.0 * nr)) {
 172                         assert(plist[j].u.f <= 100.0);
 173
 174                         if (j == oval_len) {
 175                                 oval_len += 100;
 176                                 ovals = realloc(ovals, oval_len * sizeof(unsigned int));
 177                         }
 178
 179                         ovals[j] = plat_idx_to_val(i);
 180                         if (ovals[j] < *minv)
 181                                 *minv = ovals[j];
 182                         if (ovals[j] > *maxv)
 183                                 *maxv = ovals[j];
 184
 185                         is_last = (j == len - 1);
 186                         if (is_last)
 187                                 break;
 188
 189                         j++;
 190                 }
 191         }
 192
 193         *output = ovals;
 194         return len;
 195 }
 196
 197 /*
 198  * Find and display the p-th percentile of clat
 199  */
 200 static void show_clat_percentiles(unsigned int *io_u_plat, unsigned long nr,
 201                                   fio_fp64_t *plist, unsigned int precision,
 202                                   struct buf_output *out)
 203 {
 204         unsigned int len, j = 0, minv, maxv;
 205         unsigned int *ovals;
 206         int is_last, per_line, scale_down;
 207         char fmt[32];
 208
 209         len = calc_clat_percentiles(io_u_plat, nr, plist, &ovals, &maxv, &minv);
 210         if (!len)
 211                 goto out;
 212
 213         /*
 214          * We default to usecs, but if the value range is such that we
 215          * should scale down to msecs, do that.
 216          */
 217         if (minv > 2000 && maxv > 99999) {
 218                 scale_down = 1;
 219                 log_buf(out, "    clat percentiles (msec):\n     |");
 220         } else {
 221                 scale_down = 0;
 222                 log_buf(out, "    clat percentiles (usec):\n     |");
 223         }
 224
 225         snprintf(fmt, sizeof(fmt), "%%1.%uf", precision);
 226         per_line = (80 - 7) / (precision + 14);
 227
 228         for (j = 0; j < len; j++) {
 229                 char fbuf[16], *ptr = fbuf;
 230
 231                 /* for formatting */
 232                 if (j != 0 && (j % per_line) == 0)
 233                         log_buf(out, "     |");
 234
 235                 /* end of the list */
 236                 is_last = (j == len - 1);
 237
 238                 if (plist[j].u.f < 10.0)
 239                         ptr += sprintf(fbuf, " ");
 240
 241                 snprintf(ptr, sizeof(fbuf), fmt, plist[j].u.f);
 242
 243                 if (scale_down)
 244                         ovals[j] = (ovals[j] + 999) / 1000;
 245
 246                 log_buf(out, " %sth=[%5u]%c", fbuf, ovals[j], is_last ? '\n' : ',');
 247
 248                 if (is_last)
 249                         break;
 250
 251                 if ((j % per_line) == per_line - 1)     /* for formatting */
 252                         log_buf(out, "\n");
 253         }
 254
 255 out:
 256         if (ovals)
 257                 free(ovals);
 258 }
 259
 260 bool calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 261               double *mean, double *dev)
 262 {
 263         double n = (double) is->samples;
 264
 265         if (n == 0)
 266                 return false;
 267
 268         *min = is->min_val;
 269         *max = is->max_val;
 270         *mean = is->mean.u.f;
 271
 272         if (n > 1.0)
 273                 *dev = sqrt(is->S.u.f / (n - 1.0));
 274         else
 275                 *dev = 0;
 276
 277         return true;
 278 }
 279
 280 void show_group_stats(struct group_run_stats *rs, struct buf_output *out)
 281 {
 282         char *io, *agg, *min, *max;
 283         char *ioalt, *aggalt, *minalt, *maxalt;
 284         const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
 285         int i;
 286
 287         log_buf(out, "\nRun status group %d (all jobs):\n", rs->groupid);
 288
 289         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 290                 const int i2p = is_power_of_2(rs->kb_base);
 291
 292                 if (!rs->max_run[i])
 293                         continue;
 294
 295                 io = num2str(rs->iobytes[i], 4, 1, i2p, N2S_BYTE);
 296                 ioalt = num2str(rs->iobytes[i], 4, 1, !i2p, N2S_BYTE);
 297                 agg = num2str(rs->agg[i], 4, 1, i2p, rs->unit_base);
 298                 aggalt = num2str(rs->agg[i], 4, 1, !i2p, rs->unit_base);
 299                 min = num2str(rs->min_bw[i], 4, 1, i2p, rs->unit_base);
 300                 minalt = num2str(rs->min_bw[i], 4, 1, !i2p, rs->unit_base);
 301                 max = num2str(rs->max_bw[i], 4, 1, i2p, rs->unit_base);
 302                 maxalt = num2str(rs->max_bw[i], 4, 1, !i2p, rs->unit_base);
 303                 log_buf(out, "%s: bw=%s (%s), %s-%s (%s-%s), io=%s (%s), run=%llu-%llumsec\n",
 304                                 rs->unified_rw_rep ? "  MIXED" : str[i],
 305                                 agg, aggalt, min, max, minalt, maxalt, io, ioalt,
 306                                 (unsigned long long) rs->min_run[i],
 307                                 (unsigned long long) rs->max_run[i]);
 308
 309                 free(io);
 310                 free(agg);
 311                 free(min);
 312                 free(max);
 313                 free(ioalt);
 314                 free(aggalt);
 315                 free(minalt);
 316                 free(maxalt);
 317         }
 318 }
 319
 320 void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist)
 321 {
 322         int i;
 323
 324         /*
 325          * Do depth distribution calculations
 326          */
 327         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
 328                 if (total) {
 329                         io_u_dist[i] = (double) map[i] / (double) total;
 330                         io_u_dist[i] *= 100.0;
 331                         if (io_u_dist[i] < 0.1 && map[i])
 332                                 io_u_dist[i] = 0.1;
 333                 } else
 334                         io_u_dist[i] = 0.0;
 335         }
 336 }
 337
 338 static void stat_calc_lat(struct thread_stat *ts, double *dst,
 339                           unsigned int *src, int nr)
 340 {
 341         unsigned long total = ddir_rw_sum(ts->total_io_u);
 342         int i;
 343
 344         /*
 345          * Do latency distribution calculations
 346          */
 347         for (i = 0; i < nr; i++) {
 348                 if (total) {
 349                         dst[i] = (double) src[i] / (double) total;
 350                         dst[i] *= 100.0;
 351                         if (dst[i] < 0.01 && src[i])
 352                                 dst[i] = 0.01;
 353                 } else
 354                         dst[i] = 0.0;
 355         }
 356 }
 357
 358 void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat)
 359 {
 360         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_u, FIO_IO_U_LAT_U_NR);
 361 }
 362
 363 void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat)
 364 {
 365         stat_calc_lat(ts, io_u_lat, ts->io_u_lat_m, FIO_IO_U_LAT_M_NR);
 366 }
 367
 368 static void display_lat(const char *name, unsigned long min, unsigned long max,
 369                         double mean, double dev, struct buf_output *out)
 370 {
 371         const char *base = "(usec)";
 372         char *minp, *maxp;
 373
 374         if (usec_to_msec(&min, &max, &mean, &dev))
 375                 base = "(msec)";
 376
 377         minp = num2str(min, 6, 1, 0, N2S_NONE);
 378         maxp = num2str(max, 6, 1, 0, N2S_NONE);
 379
 380         log_buf(out, "    %s %s: min=%s, max=%s, avg=%5.02f,"
 381                  " stdev=%5.02f\n", name, base, minp, maxp, mean, dev);
 382
 383         free(minp);
 384         free(maxp);
 385 }
 386
 387 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
 388                              int ddir, struct buf_output *out)
 389 {
 390         const char *str[] = { " read", "write", " trim" };
 391         unsigned long min, max, runt;
 392         unsigned long long bw, iops;
 393         double mean, dev;
 394         char *io_p, *bw_p, *bw_p_alt, *iops_p;
 395         int i2p;
 396
 397         assert(ddir_rw(ddir));
 398
 399         if (!ts->runtime[ddir])
 400                 return;
 401
 402         i2p = is_power_of_2(rs->kb_base);
 403         runt = ts->runtime[ddir];
 404
 405         bw = (1000 * ts->io_bytes[ddir]) / runt;
 406         io_p = num2str(ts->io_bytes[ddir], 4, 1, i2p, N2S_BYTE);
 407         bw_p = num2str(bw, 4, 1, i2p, ts->unit_base);
 408         bw_p_alt = num2str(bw, 4, 1, !i2p, ts->unit_base);
 409
 410         iops = (1000 * (uint64_t)ts->total_io_u[ddir]) / runt;
 411         iops_p = num2str(iops, 4, 1, 0, N2S_NONE);
 412
 413         log_buf(out, "  %s: IOPS=%s, BW=%s (%s)(%s/%llumsec)\n",
 414                         rs->unified_rw_rep ? "mixed" : str[ddir],
 415                         iops_p, bw_p, bw_p_alt, io_p,
 416                         (unsigned long long) ts->runtime[ddir]);
 417
 418         free(io_p);
 419         free(bw_p);
 420         free(bw_p_alt);
 421         free(iops_p);
 422
 423         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 424                 display_lat("slat", min, max, mean, dev, out);
 425         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 426                 display_lat("clat", min, max, mean, dev, out);
 427         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 428                 display_lat(" lat", min, max, mean, dev, out);
 429
 430         if (ts->clat_percentiles) {
 431                 show_clat_percentiles(ts->io_u_plat[ddir],
 432                                         ts->clat_stat[ddir].samples,
 433                                         ts->percentile_list,
 434                                         ts->percentile_precision, out);
 435         }
 436         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 437                 double p_of_agg = 100.0, fkb_base = (double)rs->kb_base;
 438                 const char *bw_str;
 439
 440                 if ((rs->unit_base == 1) && i2p)
 441                         bw_str = "Kibit";
 442                 else if (rs->unit_base == 1)
 443                         bw_str = "kbit";
 444                 else if (i2p)
 445                         bw_str = "KiB";
 446                 else
 447                         bw_str = "kB";
 448
 449                 if (rs->unit_base == 1) {
 450                         min *= 8.0;
 451                         max *= 8.0;
 452                         mean *= 8.0;
 453                         dev *= 8.0;
 454                 }
 455
 456                 if (rs->agg[ddir]) {
 457                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 458                         if (p_of_agg > 100.0)
 459                                 p_of_agg = 100.0;
 460                 }
 461
 462                 if (mean > fkb_base * fkb_base) {
 463                         min /= fkb_base;
 464                         max /= fkb_base;
 465                         mean /= fkb_base;
 466                         dev /= fkb_base;
 467                         bw_str = (rs->unit_base == 1 ? "Mibit" : "MiB");
 468                 }
 469
 470                 log_buf(out, "   bw (%5s/s): min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, stdev=%5.02f\n",
 471                         bw_str, min, max, p_of_agg, mean, dev);
 472         }
 473 }
 474
 475 static int show_lat(double *io_u_lat, int nr, const char **ranges,
 476                     const char *msg, struct buf_output *out)
 477 {
 478         int new_line = 1, i, line = 0, shown = 0;
 479
 480         for (i = 0; i < nr; i++) {
 481                 if (io_u_lat[i] <= 0.0)
 482                         continue;
 483                 shown = 1;
 484                 if (new_line) {
 485                         if (line)
 486                                 log_buf(out, "\n");
 487                         log_buf(out, "    lat (%s) : ", msg);
 488                         new_line = 0;
 489                         line = 0;
 490                 }
 491                 if (line)
 492                         log_buf(out, ", ");
 493                 log_buf(out, "%s%3.2f%%", ranges[i], io_u_lat[i]);
 494                 line++;
 495                 if (line == 5)
 496                         new_line = 1;
 497         }
 498
 499         if (shown)
 500                 log_buf(out, "\n");
 501
 502         return shown;
 503 }
 504
 505 static void show_lat_u(double *io_u_lat_u, struct buf_output *out)
 506 {
 507         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 508                                  "250=", "500=", "750=", "1000=", };
 509
 510         show_lat(io_u_lat_u, FIO_IO_U_LAT_U_NR, ranges, "usec", out);
 511 }
 512
 513 static void show_lat_m(double *io_u_lat_m, struct buf_output *out)
 514 {
 515         const char *ranges[] = { "2=", "4=", "10=", "20=", "50=", "100=",
 516                                  "250=", "500=", "750=", "1000=", "2000=",
 517                                  ">=2000=", };
 518
 519         show_lat(io_u_lat_m, FIO_IO_U_LAT_M_NR, ranges, "msec", out);
 520 }
 521
 522 static void show_latencies(struct thread_stat *ts, struct buf_output *out)
 523 {
 524         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
 525         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
 526
 527         stat_calc_lat_u(ts, io_u_lat_u);
 528         stat_calc_lat_m(ts, io_u_lat_m);
 529
 530         show_lat_u(io_u_lat_u, out);
 531         show_lat_m(io_u_lat_m, out);
 532 }
 533
 534 static int block_state_category(int block_state)
 535 {
 536         switch (block_state) {
 537         case BLOCK_STATE_UNINIT:
 538                 return 0;
 539         case BLOCK_STATE_TRIMMED:
 540         case BLOCK_STATE_WRITTEN:
 541                 return 1;
 542         case BLOCK_STATE_WRITE_FAILURE:
 543         case BLOCK_STATE_TRIM_FAILURE:
 544                 return 2;
 545         default:
 546                 /* Silence compile warning on some BSDs and have a return */
 547                 assert(0);
 548                 return -1;
 549         }
 550 }
 551
 552 static int compare_block_infos(const void *bs1, const void *bs2)
 553 {
 554         uint32_t block1 = *(uint32_t *)bs1;
 555         uint32_t block2 = *(uint32_t *)bs2;
 556         int state1 = BLOCK_INFO_STATE(block1);
 557         int state2 = BLOCK_INFO_STATE(block2);
 558         int bscat1 = block_state_category(state1);
 559         int bscat2 = block_state_category(state2);
 560         int cycles1 = BLOCK_INFO_TRIMS(block1);
 561         int cycles2 = BLOCK_INFO_TRIMS(block2);
 562
 563         if (bscat1 < bscat2)
 564                 return -1;
 565         if (bscat1 > bscat2)
 566                 return 1;
 567
 568         if (cycles1 < cycles2)
 569                 return -1;
 570         if (cycles1 > cycles2)
 571                 return 1;
 572
 573         if (state1 < state2)
 574                 return -1;
 575         if (state1 > state2)
 576                 return 1;
 577
 578         assert(block1 == block2);
 579         return 0;
 580 }
 581
 582 static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
 583                                   fio_fp64_t *plist, unsigned int **percentiles,
 584                                   unsigned int *types)
 585 {
 586         int len = 0;
 587         int i, nr_uninit;
 588
 589         qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
 590
 591         while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
 592                 len++;
 593
 594         if (!len)
 595                 return 0;
 596
 597         /*
 598          * Sort the percentile list. Note that it may already be sorted if
 599          * we are using the default values, but since it's a short list this
 600          * isn't a worry. Also note that this does not work for NaN values.
 601          */
 602         if (len > 1)
 603                 qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
 604
 605         nr_uninit = 0;
 606         /* Start only after the uninit entries end */
 607         for (nr_uninit = 0;
 608              nr_uninit < nr_block_infos
 609                 && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
 610              nr_uninit ++)
 611                 ;
 612
 613         if (nr_uninit == nr_block_infos)
 614                 return 0;
 615
 616         *percentiles = calloc(len, sizeof(**percentiles));
 617
 618         for (i = 0; i < len; i++) {
 619                 int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
 620                                 + nr_uninit;
 621                 (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
 622         }
 623
 624         memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
 625         for (i = 0; i < nr_block_infos; i++)
 626                 types[BLOCK_INFO_STATE(block_infos[i])]++;
 627
 628         return len;
 629 }
 630
 631 static const char *block_state_names[] = {
 632         [BLOCK_STATE_UNINIT] = "unwritten",
 633         [BLOCK_STATE_TRIMMED] = "trimmed",
 634         [BLOCK_STATE_WRITTEN] = "written",
 635         [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
 636         [BLOCK_STATE_WRITE_FAILURE] = "write failure",
 637 };
 638
 639 static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 640                              fio_fp64_t *plist, struct buf_output *out)
 641 {
 642         int len, pos, i;
 643         unsigned int *percentiles = NULL;
 644         unsigned int block_state_counts[BLOCK_STATE_COUNT];
 645
 646         len = calc_block_percentiles(nr_block_infos, block_infos, plist,
 647                                      &percentiles, block_state_counts);
 648
 649         log_buf(out, "  block lifetime percentiles :\n   |");
 650         pos = 0;
 651         for (i = 0; i < len; i++) {
 652                 uint32_t block_info = percentiles[i];
 653 #define LINE_LENGTH     75
 654                 char str[LINE_LENGTH];
 655                 int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
 656                                      plist[i].u.f, block_info,
 657                                      i == len - 1 ? '\n' : ',');
 658                 assert(strln < LINE_LENGTH);
 659                 if (pos + strln > LINE_LENGTH) {
 660                         pos = 0;
 661                         log_buf(out, "\n   |");
 662                 }
 663                 log_buf(out, "%s", str);
 664                 pos += strln;
 665 #undef LINE_LENGTH
 666         }
 667         if (percentiles)
 668                 free(percentiles);
 669
 670         log_buf(out, "        states               :");
 671         for (i = 0; i < BLOCK_STATE_COUNT; i++)
 672                 log_buf(out, " %s=%u%c",
 673                          block_state_names[i], block_state_counts[i],
 674                          i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 675 }
 676
 677 static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
 678 {
 679         char *p1, *p1alt, *p2;
 680         unsigned long long bw_mean, iops_mean;
 681         const int i2p = is_power_of_2(ts->kb_base);
 682
 683         if (!ts->ss_dur)
 684                 return;
 685
 686         bw_mean = steadystate_bw_mean(ts);
 687         iops_mean = steadystate_iops_mean(ts);
 688
 689         p1 = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, i2p, ts->unit_base);
 690         p1alt = num2str(bw_mean / ts->kb_base, 4, ts->kb_base, !i2p, ts->unit_base);
 691         p2 = num2str(iops_mean, 4, 1, 0, N2S_NONE);
 692
 693         log_buf(out, "  steadystate  : attained=%s, bw=%s (%s), iops=%s, %s%s=%.3f%s\n",
 694                 ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
 695                 p1, p1alt, p2,
 696                 ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
 697                 ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
 698                 ts->ss_criterion.u.f,
 699                 ts->ss_state & __FIO_SS_PCT ? "%" : "");
 700
 701         free(p1);
 702         free(p1alt);
 703         free(p2);
 704 }
 705
 706 static void show_thread_status_normal(struct thread_stat *ts,
 707                                       struct group_run_stats *rs,
 708                                       struct buf_output *out)
 709 {
 710         double usr_cpu, sys_cpu;
 711         unsigned long runtime;
 712         double io_u_dist[FIO_IO_U_MAP_NR];
 713         time_t time_p;
 714         char time_buf[32];
 715
 716         if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
 717                 return;
 718
 719         memset(time_buf, 0, sizeof(time_buf));
 720
 721         time(&time_p);
 722         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
 723
 724         if (!ts->error) {
 725                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d: pid=%d: %s",
 726                                         ts->name, ts->groupid, ts->members,
 727                                         ts->error, (int) ts->pid, time_buf);
 728         } else {
 729                 log_buf(out, "%s: (groupid=%d, jobs=%d): err=%2d (%s): pid=%d: %s",
 730                                         ts->name, ts->groupid, ts->members,
 731                                         ts->error, ts->verror, (int) ts->pid,
 732                                         time_buf);
 733         }
 734
 735         if (strlen(ts->description))
 736                 log_buf(out, "  Description  : [%s]\n", ts->description);
 737
 738         if (ts->io_bytes[DDIR_READ])
 739                 show_ddir_status(rs, ts, DDIR_READ, out);
 740         if (ts->io_bytes[DDIR_WRITE])
 741                 show_ddir_status(rs, ts, DDIR_WRITE, out);
 742         if (ts->io_bytes[DDIR_TRIM])
 743                 show_ddir_status(rs, ts, DDIR_TRIM, out);
 744
 745         show_latencies(ts, out);
 746
 747         runtime = ts->total_run_time;
 748         if (runtime) {
 749                 double runt = (double) runtime;
 750
 751                 usr_cpu = (double) ts->usr_time * 100 / runt;
 752                 sys_cpu = (double) ts->sys_time * 100 / runt;
 753         } else {
 754                 usr_cpu = 0;
 755                 sys_cpu = 0;
 756         }
 757
 758         log_buf(out, "  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%llu,"
 759                  " majf=%llu, minf=%llu\n", usr_cpu, sys_cpu,
 760                         (unsigned long long) ts->ctx,
 761                         (unsigned long long) ts->majf,
 762                         (unsigned long long) ts->minf);
 763
 764         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
 765         log_buf(out, "  IO depths    : 1=%3.1f%%, 2=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%,"
 766                  " 16=%3.1f%%, 32=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 767                                         io_u_dist[1], io_u_dist[2],
 768                                         io_u_dist[3], io_u_dist[4],
 769                                         io_u_dist[5], io_u_dist[6]);
 770
 771         stat_calc_dist(ts->io_u_submit, ts->total_submit, io_u_dist);
 772         log_buf(out, "     submit    : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 773                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 774                                         io_u_dist[1], io_u_dist[2],
 775                                         io_u_dist[3], io_u_dist[4],
 776                                         io_u_dist[5], io_u_dist[6]);
 777         stat_calc_dist(ts->io_u_complete, ts->total_complete, io_u_dist);
 778         log_buf(out, "     complete  : 0=%3.1f%%, 4=%3.1f%%, 8=%3.1f%%, 16=%3.1f%%,"
 779                  " 32=%3.1f%%, 64=%3.1f%%, >=64=%3.1f%%\n", io_u_dist[0],
 780                                         io_u_dist[1], io_u_dist[2],
 781                                         io_u_dist[3], io_u_dist[4],
 782                                         io_u_dist[5], io_u_dist[6]);
 783         log_buf(out, "     issued rwt: total=%llu,%llu,%llu,"
 784                                  " short=%llu,%llu,%llu,"
 785                                  " dropped=%llu,%llu,%llu\n",
 786                                         (unsigned long long) ts->total_io_u[0],
 787                                         (unsigned long long) ts->total_io_u[1],
 788                                         (unsigned long long) ts->total_io_u[2],
 789                                         (unsigned long long) ts->short_io_u[0],
 790                                         (unsigned long long) ts->short_io_u[1],
 791                                         (unsigned long long) ts->short_io_u[2],
 792                                         (unsigned long long) ts->drop_io_u[0],
 793                                         (unsigned long long) ts->drop_io_u[1],
 794                                         (unsigned long long) ts->drop_io_u[2]);
 795         if (ts->continue_on_error) {
 796                 log_buf(out, "     errors    : total=%llu, first_error=%d/<%s>\n",
 797                                         (unsigned long long)ts->total_err_count,
 798                                         ts->first_error,
 799                                         strerror(ts->first_error));
 800         }
 801         if (ts->latency_depth) {
 802                 log_buf(out, "     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
 803                                         (unsigned long long)ts->latency_target,
 804                                         (unsigned long long)ts->latency_window,
 805                                         ts->latency_percentile.u.f,
 806                                         ts->latency_depth);
 807         }
 808
 809         if (ts->nr_block_infos)
 810                 show_block_infos(ts->nr_block_infos, ts->block_infos,
 811                                   ts->percentile_list, out);
 812
 813         if (ts->ss_dur)
 814                 show_ss_normal(ts, out);
 815 }
 816
 817 static void show_ddir_status_terse(struct thread_stat *ts,
 818                                    struct group_run_stats *rs, int ddir,
 819                                    struct buf_output *out)
 820 {
 821         unsigned long min, max;
 822         unsigned long long bw, iops;
 823         unsigned int *ovals = NULL;
 824         double mean, dev;
 825         unsigned int len, minv, maxv;
 826         int i;
 827
 828         assert(ddir_rw(ddir));
 829
 830         iops = bw = 0;
 831         if (ts->runtime[ddir]) {
 832                 uint64_t runt = ts->runtime[ddir];
 833
 834                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 835                 iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
 836         }
 837
 838         log_buf(out, ";%llu;%llu;%llu;%llu",
 839                 (unsigned long long) ts->io_bytes[ddir] >> 10, bw, iops,
 840                                         (unsigned long long) ts->runtime[ddir]);
 841
 842         if (calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev))
 843                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 844         else
 845                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 846
 847         if (calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev))
 848                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 849         else
 850                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 851
 852         if (ts->clat_percentiles) {
 853                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 854                                         ts->clat_stat[ddir].samples,
 855                                         ts->percentile_list, &ovals, &maxv,
 856                                         &minv);
 857         } else
 858                 len = 0;
 859
 860         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 861                 if (i >= len) {
 862                         log_buf(out, ";0%%=0");
 863                         continue;
 864                 }
 865                 log_buf(out, ";%f%%=%u", ts->percentile_list[i].u.f, ovals[i]);
 866         }
 867
 868         if (calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev))
 869                 log_buf(out, ";%lu;%lu;%f;%f", min, max, mean, dev);
 870         else
 871                 log_buf(out, ";%lu;%lu;%f;%f", 0UL, 0UL, 0.0, 0.0);
 872
 873         if (ovals)
 874                 free(ovals);
 875
 876         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 877                 double p_of_agg = 100.0;
 878
 879                 if (rs->agg[ddir]) {
 880                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 881                         if (p_of_agg > 100.0)
 882                                 p_of_agg = 100.0;
 883                 }
 884
 885                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", min, max, p_of_agg, mean, dev);
 886         } else
 887                 log_buf(out, ";%lu;%lu;%f%%;%f;%f", 0UL, 0UL, 0.0, 0.0, 0.0);
 888 }
 889
 890 static void add_ddir_status_json(struct thread_stat *ts,
 891                 struct group_run_stats *rs, int ddir, struct json_object *parent)
 892 {
 893         unsigned long min, max;
 894         unsigned long long bw;
 895         unsigned int *ovals = NULL;
 896         double mean, dev, iops;
 897         unsigned int len, minv, maxv;
 898         int i;
 899         const char *ddirname[] = {"read", "write", "trim"};
 900         struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
 901         char buf[120];
 902         double p_of_agg = 100.0;
 903
 904         assert(ddir_rw(ddir));
 905
 906         if (ts->unified_rw_rep && ddir != DDIR_READ)
 907                 return;
 908
 909         dir_object = json_create_object();
 910         json_object_add_value_object(parent,
 911                 ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 912
 913         bw = 0;
 914         iops = 0.0;
 915         if (ts->runtime[ddir]) {
 916                 uint64_t runt = ts->runtime[ddir];
 917
 918                 bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024; /* KiB/s */
 919                 iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
 920         }
 921
 922         json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
 923         json_object_add_value_int(dir_object, "bw", bw);
 924         json_object_add_value_float(dir_object, "iops", iops);
 925         json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
 926         json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
 927         json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
 928         json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 929
 930         if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
 931                 min = max = 0;
 932                 mean = dev = 0.0;
 933         }
 934         tmp_object = json_create_object();
 935         json_object_add_value_object(dir_object, "slat", tmp_object);
 936         json_object_add_value_int(tmp_object, "min", min);
 937         json_object_add_value_int(tmp_object, "max", max);
 938         json_object_add_value_float(tmp_object, "mean", mean);
 939         json_object_add_value_float(tmp_object, "stddev", dev);
 940
 941         if (!calc_lat(&ts->clat_stat[ddir], &min, &max, &mean, &dev)) {
 942                 min = max = 0;
 943                 mean = dev = 0.0;
 944         }
 945         tmp_object = json_create_object();
 946         json_object_add_value_object(dir_object, "clat", tmp_object);
 947         json_object_add_value_int(tmp_object, "min", min);
 948         json_object_add_value_int(tmp_object, "max", max);
 949         json_object_add_value_float(tmp_object, "mean", mean);
 950         json_object_add_value_float(tmp_object, "stddev", dev);
 951
 952         if (ts->clat_percentiles) {
 953                 len = calc_clat_percentiles(ts->io_u_plat[ddir],
 954                                         ts->clat_stat[ddir].samples,
 955                                         ts->percentile_list, &ovals, &maxv,
 956                                         &minv);
 957         } else
 958                 len = 0;
 959
 960         percentile_object = json_create_object();
 961         json_object_add_value_object(tmp_object, "percentile", percentile_object);
 962         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 963                 if (i >= len) {
 964                         json_object_add_value_int(percentile_object, "0.00", 0);
 965                         continue;
 966                 }
 967                 snprintf(buf, sizeof(buf), "%f", ts->percentile_list[i].u.f);
 968                 json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
 969         }
 970
 971         if (output_format & FIO_OUTPUT_JSON_PLUS) {
 972                 clat_bins_object = json_create_object();
 973                 json_object_add_value_object(tmp_object, "bins", clat_bins_object);
 974                 for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
 975                         if (ts->io_u_plat[ddir][i]) {
 976                                 snprintf(buf, sizeof(buf), "%llu", plat_idx_to_val(i));
 977                                 json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
 978                         }
 979                 }
 980         }
 981
 982         if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
 983                 min = max = 0;
 984                 mean = dev = 0.0;
 985         }
 986         tmp_object = json_create_object();
 987         json_object_add_value_object(dir_object, "lat", tmp_object);
 988         json_object_add_value_int(tmp_object, "min", min);
 989         json_object_add_value_int(tmp_object, "max", max);
 990         json_object_add_value_float(tmp_object, "mean", mean);
 991         json_object_add_value_float(tmp_object, "stddev", dev);
 992         if (ovals)
 993                 free(ovals);
 994
 995         if (calc_lat(&ts->bw_stat[ddir], &min, &max, &mean, &dev)) {
 996                 if (rs->agg[ddir]) {
 997                         p_of_agg = mean * 100 / (double) rs->agg[ddir];
 998                         if (p_of_agg > 100.0)
 999                                 p_of_agg = 100.0;
1000                 }
1001         } else {
1002                 min = max = 0;
1003                 p_of_agg = mean = dev = 0.0;
1004         }
1005         json_object_add_value_int(dir_object, "bw_min", min);
1006         json_object_add_value_int(dir_object, "bw_max", max);
1007         json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
1008         json_object_add_value_float(dir_object, "bw_mean", mean);
1009         json_object_add_value_float(dir_object, "bw_dev", dev);
1010 }
1011
1012 static void show_thread_status_terse_v2(struct thread_stat *ts,
1013                                         struct group_run_stats *rs,
1014                                         struct buf_output *out)
1015 {
1016         double io_u_dist[FIO_IO_U_MAP_NR];
1017         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1018         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1019         double usr_cpu, sys_cpu;
1020         int i;
1021
1022         /* General Info */
1023         log_buf(out, "2;%s;%d;%d", ts->name, ts->groupid, ts->error);
1024         /* Log Read Status */
1025         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1026         /* Log Write Status */
1027         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1028         /* Log Trim Status */
1029         show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1030
1031         /* CPU Usage */
1032         if (ts->total_run_time) {
1033                 double runt = (double) ts->total_run_time;
1034
1035                 usr_cpu = (double) ts->usr_time * 100 / runt;
1036                 sys_cpu = (double) ts->sys_time * 100 / runt;
1037         } else {
1038                 usr_cpu = 0;
1039                 sys_cpu = 0;
1040         }
1041
1042         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1043                                                 (unsigned long long) ts->ctx,
1044                                                 (unsigned long long) ts->majf,
1045                                                 (unsigned long long) ts->minf);
1046
1047         /* Calc % distribution of IO depths, usecond, msecond latency */
1048         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1049         stat_calc_lat_u(ts, io_u_lat_u);
1050         stat_calc_lat_m(ts, io_u_lat_m);
1051
1052         /* Only show fixed 7 I/O depth levels*/
1053         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1054                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1055                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1056
1057         /* Microsecond latency */
1058         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1059                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1060         /* Millisecond latency */
1061         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1062                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1063         /* Additional output if continue_on_error set - default off*/
1064         if (ts->continue_on_error)
1065                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1066         log_buf(out, "\n");
1067
1068         /* Additional output if description is set */
1069         if (strlen(ts->description))
1070                 log_buf(out, ";%s", ts->description);
1071
1072         log_buf(out, "\n");
1073 }
1074
1075 static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
1076                                            struct group_run_stats *rs, int ver,
1077                                            struct buf_output *out)
1078 {
1079         double io_u_dist[FIO_IO_U_MAP_NR];
1080         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1081         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1082         double usr_cpu, sys_cpu;
1083         int i;
1084
1085         /* General Info */
1086         log_buf(out, "%d;%s;%s;%d;%d", ver, fio_version_string,
1087                                         ts->name, ts->groupid, ts->error);
1088         /* Log Read Status */
1089         show_ddir_status_terse(ts, rs, DDIR_READ, out);
1090         /* Log Write Status */
1091         show_ddir_status_terse(ts, rs, DDIR_WRITE, out);
1092         /* Log Trim Status */
1093         if (ver == 4)
1094                 show_ddir_status_terse(ts, rs, DDIR_TRIM, out);
1095
1096         /* CPU Usage */
1097         if (ts->total_run_time) {
1098                 double runt = (double) ts->total_run_time;
1099
1100                 usr_cpu = (double) ts->usr_time * 100 / runt;
1101                 sys_cpu = (double) ts->sys_time * 100 / runt;
1102         } else {
1103                 usr_cpu = 0;
1104                 sys_cpu = 0;
1105         }
1106
1107         log_buf(out, ";%f%%;%f%%;%llu;%llu;%llu", usr_cpu, sys_cpu,
1108                                                 (unsigned long long) ts->ctx,
1109                                                 (unsigned long long) ts->majf,
1110                                                 (unsigned long long) ts->minf);
1111
1112         /* Calc % distribution of IO depths, usecond, msecond latency */
1113         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1114         stat_calc_lat_u(ts, io_u_lat_u);
1115         stat_calc_lat_m(ts, io_u_lat_m);
1116
1117         /* Only show fixed 7 I/O depth levels*/
1118         log_buf(out, ";%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%;%3.1f%%",
1119                         io_u_dist[0], io_u_dist[1], io_u_dist[2], io_u_dist[3],
1120                         io_u_dist[4], io_u_dist[5], io_u_dist[6]);
1121
1122         /* Microsecond latency */
1123         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1124                 log_buf(out, ";%3.2f%%", io_u_lat_u[i]);
1125         /* Millisecond latency */
1126         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1127                 log_buf(out, ";%3.2f%%", io_u_lat_m[i]);
1128
1129         /* disk util stats, if any */
1130         show_disk_util(1, NULL, out);
1131
1132         /* Additional output if continue_on_error set - default off*/
1133         if (ts->continue_on_error)
1134                 log_buf(out, ";%llu;%d", (unsigned long long) ts->total_err_count, ts->first_error);
1135
1136         /* Additional output if description is set */
1137         if (strlen(ts->description))
1138                 log_buf(out, ";%s", ts->description);
1139
1140         log_buf(out, "\n");
1141 }
1142
1143 static void json_add_job_opts(struct json_object *root, const char *name,
1144                               struct flist_head *opt_list, bool num_jobs)
1145 {
1146         struct json_object *dir_object;
1147         struct flist_head *entry;
1148         struct print_option *p;
1149
1150         if (flist_empty(opt_list))
1151                 return;
1152
1153         dir_object = json_create_object();
1154         json_object_add_value_object(root, name, dir_object);
1155
1156         flist_for_each(entry, opt_list) {
1157                 const char *pos = "";
1158
1159                 p = flist_entry(entry, struct print_option, list);
1160                 if (!num_jobs && !strcmp(p->name, "numjobs"))
1161                         continue;
1162                 if (p->value)
1163                         pos = p->value;
1164                 json_object_add_value_string(dir_object, p->name, pos);
1165         }
1166 }
1167
1168 static struct json_object *show_thread_status_json(struct thread_stat *ts,
1169                                                    struct group_run_stats *rs,
1170                                                    struct flist_head *opt_list)
1171 {
1172         struct json_object *root, *tmp;
1173         struct jobs_eta *je;
1174         double io_u_dist[FIO_IO_U_MAP_NR];
1175         double io_u_lat_u[FIO_IO_U_LAT_U_NR];
1176         double io_u_lat_m[FIO_IO_U_LAT_M_NR];
1177         double usr_cpu, sys_cpu;
1178         int i;
1179         size_t size;
1180
1181         root = json_create_object();
1182         json_object_add_value_string(root, "jobname", ts->name);
1183         json_object_add_value_int(root, "groupid", ts->groupid);
1184         json_object_add_value_int(root, "error", ts->error);
1185
1186         /* ETA Info */
1187         je = get_jobs_eta(true, &size);
1188         if (je) {
1189                 json_object_add_value_int(root, "eta", je->eta_sec);
1190                 json_object_add_value_int(root, "elapsed", je->elapsed_sec);
1191         }
1192
1193         if (opt_list)
1194                 json_add_job_opts(root, "job options", opt_list, true);
1195
1196         add_ddir_status_json(ts, rs, DDIR_READ, root);
1197         add_ddir_status_json(ts, rs, DDIR_WRITE, root);
1198         add_ddir_status_json(ts, rs, DDIR_TRIM, root);
1199
1200         /* CPU Usage */
1201         if (ts->total_run_time) {
1202                 double runt = (double) ts->total_run_time;
1203
1204                 usr_cpu = (double) ts->usr_time * 100 / runt;
1205                 sys_cpu = (double) ts->sys_time * 100 / runt;
1206         } else {
1207                 usr_cpu = 0;
1208                 sys_cpu = 0;
1209         }
1210         json_object_add_value_float(root, "usr_cpu", usr_cpu);
1211         json_object_add_value_float(root, "sys_cpu", sys_cpu);
1212         json_object_add_value_int(root, "ctx", ts->ctx);
1213         json_object_add_value_int(root, "majf", ts->majf);
1214         json_object_add_value_int(root, "minf", ts->minf);
1215
1216
1217         /* Calc % distribution of IO depths, usecond, msecond latency */
1218         stat_calc_dist(ts->io_u_map, ddir_rw_sum(ts->total_io_u), io_u_dist);
1219         stat_calc_lat_u(ts, io_u_lat_u);
1220         stat_calc_lat_m(ts, io_u_lat_m);
1221
1222         tmp = json_create_object();
1223         json_object_add_value_object(root, "iodepth_level", tmp);
1224         /* Only show fixed 7 I/O depth levels*/
1225         for (i = 0; i < 7; i++) {
1226                 char name[20];
1227                 if (i < 6)
1228                         snprintf(name, 20, "%d", 1 << i);
1229                 else
1230                         snprintf(name, 20, ">=%d", 1 << i);
1231                 json_object_add_value_float(tmp, (const char *)name, io_u_dist[i]);
1232         }
1233
1234         tmp = json_create_object();
1235         json_object_add_value_object(root, "latency_us", tmp);
1236         /* Microsecond latency */
1237         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
1238                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1239                                  "250", "500", "750", "1000", };
1240                 json_object_add_value_float(tmp, ranges[i], io_u_lat_u[i]);
1241         }
1242         /* Millisecond latency */
1243         tmp = json_create_object();
1244         json_object_add_value_object(root, "latency_ms", tmp);
1245         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++) {
1246                 const char *ranges[] = { "2", "4", "10", "20", "50", "100",
1247                                  "250", "500", "750", "1000", "2000",
1248                                  ">=2000", };
1249                 json_object_add_value_float(tmp, ranges[i], io_u_lat_m[i]);
1250         }
1251
1252         /* Additional output if continue_on_error set - default off*/
1253         if (ts->continue_on_error) {
1254                 json_object_add_value_int(root, "total_err", ts->total_err_count);
1255                 json_object_add_value_int(root, "first_error", ts->first_error);
1256         }
1257
1258         if (ts->latency_depth) {
1259                 json_object_add_value_int(root, "latency_depth", ts->latency_depth);
1260                 json_object_add_value_int(root, "latency_target", ts->latency_target);
1261                 json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
1262                 json_object_add_value_int(root, "latency_window", ts->latency_window);
1263         }
1264
1265         /* Additional output if description is set */
1266         if (strlen(ts->description))
1267                 json_object_add_value_string(root, "desc", ts->description);
1268
1269         if (ts->nr_block_infos) {
1270                 /* Block error histogram and types */
1271                 int len;
1272                 unsigned int *percentiles = NULL;
1273                 unsigned int block_state_counts[BLOCK_STATE_COUNT];
1274
1275                 len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
1276                                              ts->percentile_list,
1277                                              &percentiles, block_state_counts);
1278
1279                 if (len) {
1280                         struct json_object *block, *percentile_object, *states;
1281                         int state;
1282                         block = json_create_object();
1283                         json_object_add_value_object(root, "block", block);
1284
1285                         percentile_object = json_create_object();
1286                         json_object_add_value_object(block, "percentiles",
1287                                                      percentile_object);
1288                         for (i = 0; i < len; i++) {
1289                                 char buf[20];
1290                                 snprintf(buf, sizeof(buf), "%f",
1291                                          ts->percentile_list[i].u.f);
1292                                 json_object_add_value_int(percentile_object,
1293                                                           (const char *)buf,
1294                                                           percentiles[i]);
1295                         }
1296
1297                         states = json_create_object();
1298                         json_object_add_value_object(block, "states", states);
1299                         for (state = 0; state < BLOCK_STATE_COUNT; state++) {
1300                                 json_object_add_value_int(states,
1301                                         block_state_names[state],
1302                                         block_state_counts[state]);
1303                         }
1304                         free(percentiles);
1305                 }
1306         }
1307
1308         if (ts->ss_dur) {
1309                 struct json_object *data;
1310                 struct json_array *iops, *bw;
1311                 int i, j, k;
1312                 char ss_buf[64];
1313
1314                 snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
1315                         ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
1316                         ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
1317                         (float) ts->ss_limit.u.f,
1318                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1319
1320                 tmp = json_create_object();
1321                 json_object_add_value_object(root, "steadystate", tmp);
1322                 json_object_add_value_string(tmp, "ss", ss_buf);
1323                 json_object_add_value_int(tmp, "duration", (int)ts->ss_dur);
1324                 json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
1325
1326                 snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
1327                         ts->ss_state & __FIO_SS_PCT ? "%" : "");
1328                 json_object_add_value_string(tmp, "criterion", ss_buf);
1329                 json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
1330                 json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
1331
1332                 data = json_create_object();
1333                 json_object_add_value_object(tmp, "data", data);
1334                 bw = json_create_array();
1335                 iops = json_create_array();
1336
1337                 /*
1338                 ** if ss was attained or the buffer is not full,
1339                 ** ss->head points to the first element in the list.
1340                 ** otherwise it actually points to the second element
1341                 ** in the list
1342                 */
1343                 if ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
1344                         j = ts->ss_head;
1345                 else
1346                         j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
1347                 for (i = 0; i < ts->ss_dur; i++) {
1348                         k = (j + i) % ts->ss_dur;
1349                         json_array_add_value_int(bw, ts->ss_bw_data[k]);
1350                         json_array_add_value_int(iops, ts->ss_iops_data[k]);
1351                 }
1352                 json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
1353                 json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
1354                 json_object_add_value_array(data, "iops", iops);
1355                 json_object_add_value_array(data, "bw", bw);
1356         }
1357
1358         return root;
1359 }
1360
1361 static void show_thread_status_terse(struct thread_stat *ts,
1362                                      struct group_run_stats *rs,
1363                                      struct buf_output *out)
1364 {
1365         if (terse_version == 2)
1366                 show_thread_status_terse_v2(ts, rs, out);
1367         else if (terse_version == 3 || terse_version == 4)
1368                 show_thread_status_terse_v3_v4(ts, rs, terse_version, out);
1369         else
1370                 log_err("fio: bad terse version!? %d\n", terse_version);
1371 }
1372
1373 struct json_object *show_thread_status(struct thread_stat *ts,
1374                                        struct group_run_stats *rs,
1375                                        struct flist_head *opt_list,
1376                                        struct buf_output *out)
1377 {
1378         struct json_object *ret = NULL;
1379
1380         if (output_format & FIO_OUTPUT_TERSE)
1381                 show_thread_status_terse(ts, rs,  out);
1382         if (output_format & FIO_OUTPUT_JSON)
1383                 ret = show_thread_status_json(ts, rs, opt_list);
1384         if (output_format & FIO_OUTPUT_NORMAL)
1385                 show_thread_status_normal(ts, rs,  out);
1386
1387         return ret;
1388 }
1389
1390 static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
1391 {
1392         double mean, S;
1393
1394         if (src->samples == 0)
1395                 return;
1396
1397         dst->min_val = min(dst->min_val, src->min_val);
1398         dst->max_val = max(dst->max_val, src->max_val);
1399
1400         /*
1401          * Compute new mean and S after the merge
1402          * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
1403          *  #Parallel_algorithm>
1404          */
1405         if (first) {
1406                 mean = src->mean.u.f;
1407                 S = src->S.u.f;
1408         } else {
1409                 double delta = src->mean.u.f - dst->mean.u.f;
1410
1411                 mean = ((src->mean.u.f * src->samples) +
1412                         (dst->mean.u.f * dst->samples)) /
1413                         (dst->samples + src->samples);
1414
1415                 S =  src->S.u.f + dst->S.u.f + pow(delta, 2.0) *
1416                         (dst->samples * src->samples) /
1417                         (dst->samples + src->samples);
1418         }
1419
1420         dst->samples += src->samples;
1421         dst->mean.u.f = mean;
1422         dst->S.u.f = S;
1423 }
1424
1425 void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
1426 {
1427         int i;
1428
1429         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1430                 if (dst->max_run[i] < src->max_run[i])
1431                         dst->max_run[i] = src->max_run[i];
1432                 if (dst->min_run[i] && dst->min_run[i] > src->min_run[i])
1433                         dst->min_run[i] = src->min_run[i];
1434                 if (dst->max_bw[i] < src->max_bw[i])
1435                         dst->max_bw[i] = src->max_bw[i];
1436                 if (dst->min_bw[i] && dst->min_bw[i] > src->min_bw[i])
1437                         dst->min_bw[i] = src->min_bw[i];
1438
1439                 dst->iobytes[i] += src->iobytes[i];
1440                 dst->agg[i] += src->agg[i];
1441         }
1442
1443         if (!dst->kb_base)
1444                 dst->kb_base = src->kb_base;
1445         if (!dst->unit_base)
1446                 dst->unit_base = src->unit_base;
1447 }
1448
1449 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
1450                       bool first)
1451 {
1452         int l, k;
1453
1454         for (l = 0; l < DDIR_RWDIR_CNT; l++) {
1455                 if (!dst->unified_rw_rep) {
1456                         sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
1457                         sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
1458                         sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
1459                         sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
1460
1461                         dst->io_bytes[l] += src->io_bytes[l];
1462
1463                         if (dst->runtime[l] < src->runtime[l])
1464                                 dst->runtime[l] = src->runtime[l];
1465                 } else {
1466                         sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
1467                         sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
1468                         sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
1469                         sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
1470
1471                         dst->io_bytes[0] += src->io_bytes[l];
1472
1473                         if (dst->runtime[0] < src->runtime[l])
1474                                 dst->runtime[0] = src->runtime[l];
1475
1476                         /*
1477                          * We're summing to the same destination, so override
1478                          * 'first' after the first iteration of the loop
1479                          */
1480                         first = false;
1481                 }
1482         }
1483
1484         dst->usr_time += src->usr_time;
1485         dst->sys_time += src->sys_time;
1486         dst->ctx += src->ctx;
1487         dst->majf += src->majf;
1488         dst->minf += src->minf;
1489
1490         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1491                 dst->io_u_map[k] += src->io_u_map[k];
1492         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1493                 dst->io_u_submit[k] += src->io_u_submit[k];
1494         for (k = 0; k < FIO_IO_U_MAP_NR; k++)
1495                 dst->io_u_complete[k] += src->io_u_complete[k];
1496         for (k = 0; k < FIO_IO_U_LAT_U_NR; k++)
1497                 dst->io_u_lat_u[k] += src->io_u_lat_u[k];
1498         for (k = 0; k < FIO_IO_U_LAT_M_NR; k++)
1499                 dst->io_u_lat_m[k] += src->io_u_lat_m[k];
1500
1501         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1502                 if (!dst->unified_rw_rep) {
1503                         dst->total_io_u[k] += src->total_io_u[k];
1504                         dst->short_io_u[k] += src->short_io_u[k];
1505                         dst->drop_io_u[k] += src->drop_io_u[k];
1506                 } else {
1507                         dst->total_io_u[0] += src->total_io_u[k];
1508                         dst->short_io_u[0] += src->short_io_u[k];
1509                         dst->drop_io_u[0] += src->drop_io_u[k];
1510                 }
1511         }
1512
1513         for (k = 0; k < DDIR_RWDIR_CNT; k++) {
1514                 int m;
1515
1516                 for (m = 0; m < FIO_IO_U_PLAT_NR; m++) {
1517                         if (!dst->unified_rw_rep)
1518                                 dst->io_u_plat[k][m] += src->io_u_plat[k][m];
1519                         else
1520                                 dst->io_u_plat[0][m] += src->io_u_plat[k][m];
1521                 }
1522         }
1523
1524         dst->total_run_time += src->total_run_time;
1525         dst->total_submit += src->total_submit;
1526         dst->total_complete += src->total_complete;
1527 }
1528
1529 void init_group_run_stat(struct group_run_stats *gs)
1530 {
1531         int i;
1532         memset(gs, 0, sizeof(*gs));
1533
1534         for (i = 0; i < DDIR_RWDIR_CNT; i++)
1535                 gs->min_bw[i] = gs->min_run[i] = ~0UL;
1536 }
1537
1538 void init_thread_stat(struct thread_stat *ts)
1539 {
1540         int j;
1541
1542         memset(ts, 0, sizeof(*ts));
1543
1544         for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1545                 ts->lat_stat[j].min_val = -1UL;
1546                 ts->clat_stat[j].min_val = -1UL;
1547                 ts->slat_stat[j].min_val = -1UL;
1548                 ts->bw_stat[j].min_val = -1UL;
1549         }
1550         ts->groupid = -1;
1551 }
1552
1553 void __show_run_stats(void)
1554 {
1555         struct group_run_stats *runstats, *rs;
1556         struct thread_data *td;
1557         struct thread_stat *threadstats, *ts;
1558         int i, j, k, nr_ts, last_ts, idx;
1559         int kb_base_warned = 0;
1560         int unit_base_warned = 0;
1561         struct json_object *root = NULL;
1562         struct json_array *array = NULL;
1563         struct buf_output output[FIO_OUTPUT_NR];
1564         struct flist_head **opt_lists;
1565
1566         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
1567
1568         for (i = 0; i < groupid + 1; i++)
1569                 init_group_run_stat(&runstats[i]);
1570
1571         /*
1572          * find out how many threads stats we need. if group reporting isn't
1573          * enabled, it's one-per-td.
1574          */
1575         nr_ts = 0;
1576         last_ts = -1;
1577         for_each_td(td, i) {
1578                 if (!td->o.group_reporting) {
1579                         nr_ts++;
1580                         continue;
1581                 }
1582                 if (last_ts == td->groupid)
1583                         continue;
1584                 if (!td->o.stats)
1585                         continue;
1586
1587                 last_ts = td->groupid;
1588                 nr_ts++;
1589         }
1590
1591         threadstats = malloc(nr_ts * sizeof(struct thread_stat));
1592         opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
1593
1594         for (i = 0; i < nr_ts; i++) {
1595                 init_thread_stat(&threadstats[i]);
1596                 opt_lists[i] = NULL;
1597         }
1598
1599         j = 0;
1600         last_ts = -1;
1601         idx = 0;
1602         for_each_td(td, i) {
1603                 if (!td->o.stats)
1604                         continue;
1605                 if (idx && (!td->o.group_reporting ||
1606                     (td->o.group_reporting && last_ts != td->groupid))) {
1607                         idx = 0;
1608                         j++;
1609                 }
1610
1611                 last_ts = td->groupid;
1612
1613                 ts = &threadstats[j];
1614
1615                 ts->clat_percentiles = td->o.clat_percentiles;
1616                 ts->percentile_precision = td->o.percentile_precision;
1617                 memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
1618                 opt_lists[j] = &td->opt_list;
1619
1620                 idx++;
1621                 ts->members++;
1622
1623                 if (ts->groupid == -1) {
1624                         /*
1625                          * These are per-group shared already
1626                          */
1627                         strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
1628                         if (td->o.description)
1629                                 strncpy(ts->description, td->o.description,
1630                                                 FIO_JOBDESC_SIZE - 1);
1631                         else
1632                                 memset(ts->description, 0, FIO_JOBDESC_SIZE);
1633
1634                         /*
1635                          * If multiple entries in this group, this is
1636                          * the first member.
1637                          */
1638                         ts->thread_number = td->thread_number;
1639                         ts->groupid = td->groupid;
1640
1641                         /*
1642                          * first pid in group, not very useful...
1643                          */
1644                         ts->pid = td->pid;
1645
1646                         ts->kb_base = td->o.kb_base;
1647                         ts->unit_base = td->o.unit_base;
1648                         ts->unified_rw_rep = td->o.unified_rw_rep;
1649                 } else if (ts->kb_base != td->o.kb_base && !kb_base_warned) {
1650                         log_info("fio: kb_base differs for jobs in group, using"
1651                                  " %u as the base\n", ts->kb_base);
1652                         kb_base_warned = 1;
1653                 } else if (ts->unit_base != td->o.unit_base && !unit_base_warned) {
1654                         log_info("fio: unit_base differs for jobs in group, using"
1655                                  " %u as the base\n", ts->unit_base);
1656                         unit_base_warned = 1;
1657                 }
1658
1659                 ts->continue_on_error = td->o.continue_on_error;
1660                 ts->total_err_count += td->total_err_count;
1661                 ts->first_error = td->first_error;
1662                 if (!ts->error) {
1663                         if (!td->error && td->o.continue_on_error &&
1664                             td->first_error) {
1665                                 ts->error = td->first_error;
1666                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1667                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1668                         } else  if (td->error) {
1669                                 ts->error = td->error;
1670                                 ts->verror[sizeof(ts->verror) - 1] = '\0';
1671                                 strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
1672                         }
1673                 }
1674
1675                 ts->latency_depth = td->latency_qd;
1676                 ts->latency_target = td->o.latency_target;
1677                 ts->latency_percentile = td->o.latency_percentile;
1678                 ts->latency_window = td->o.latency_window;
1679
1680                 ts->nr_block_infos = td->ts.nr_block_infos;
1681                 for (k = 0; k < ts->nr_block_infos; k++)
1682                         ts->block_infos[k] = td->ts.block_infos[k];
1683
1684                 sum_thread_stats(ts, &td->ts, idx == 1);
1685
1686                 if (td->o.ss_dur) {
1687                         ts->ss_state = td->ss.state;
1688                         ts->ss_dur = td->ss.dur;
1689                         ts->ss_head = td->ss.head;
1690                         ts->ss_bw_data = td->ss.bw_data;
1691                         ts->ss_iops_data = td->ss.iops_data;
1692                         ts->ss_limit.u.f = td->ss.limit;
1693                         ts->ss_slope.u.f = td->ss.slope;
1694                         ts->ss_deviation.u.f = td->ss.deviation;
1695                         ts->ss_criterion.u.f = td->ss.criterion;
1696                 }
1697                 else
1698                         ts->ss_dur = ts->ss_state = 0;
1699         }
1700
1701         for (i = 0; i < nr_ts; i++) {
1702                 unsigned long long bw;
1703
1704                 ts = &threadstats[i];
1705                 if (ts->groupid == -1)
1706                         continue;
1707                 rs = &runstats[ts->groupid];
1708                 rs->kb_base = ts->kb_base;
1709                 rs->unit_base = ts->unit_base;
1710                 rs->unified_rw_rep += ts->unified_rw_rep;
1711
1712                 for (j = 0; j < DDIR_RWDIR_CNT; j++) {
1713                         if (!ts->runtime[j])
1714                                 continue;
1715                         if (ts->runtime[j] < rs->min_run[j] || !rs->min_run[j])
1716                                 rs->min_run[j] = ts->runtime[j];
1717                         if (ts->runtime[j] > rs->max_run[j])
1718                                 rs->max_run[j] = ts->runtime[j];
1719
1720                         bw = 0;
1721                         if (ts->runtime[j])
1722                                 bw = ts->io_bytes[j] * 1000 / ts->runtime[j];
1723                         if (bw < rs->min_bw[j])
1724                                 rs->min_bw[j] = bw;
1725                         if (bw > rs->max_bw[j])
1726                                 rs->max_bw[j] = bw;
1727
1728                         rs->iobytes[j] += ts->io_bytes[j];
1729                 }
1730         }
1731
1732         for (i = 0; i < groupid + 1; i++) {
1733                 int ddir;
1734
1735                 rs = &runstats[i];
1736
1737                 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
1738                         if (rs->max_run[ddir])
1739                                 rs->agg[ddir] = (rs->iobytes[ddir] * 1000) /
1740                                                 rs->max_run[ddir];
1741                 }
1742         }
1743
1744         for (i = 0; i < FIO_OUTPUT_NR; i++)
1745                 buf_output_init(&output[i]);
1746
1747         /*
1748          * don't overwrite last signal output
1749          */
1750         if (output_format & FIO_OUTPUT_NORMAL)
1751                 log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
1752         if (output_format & FIO_OUTPUT_JSON) {
1753                 struct thread_data *global;
1754                 char time_buf[32];
1755                 struct timeval now;
1756                 unsigned long long ms_since_epoch;
1757
1758                 gettimeofday(&now, NULL);
1759                 ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
1760                                  (unsigned long long)(now.tv_usec) / 1000;
1761
1762                 os_ctime_r((const time_t *) &now.tv_sec, time_buf,
1763                                 sizeof(time_buf));
1764                 if (time_buf[strlen(time_buf) - 1] == '\n')
1765                         time_buf[strlen(time_buf) - 1] = '\0';
1766
1767                 root = json_create_object();
1768                 json_object_add_value_string(root, "fio version", fio_version_string);
1769                 json_object_add_value_int(root, "timestamp", now.tv_sec);
1770                 json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
1771                 json_object_add_value_string(root, "time", time_buf);
1772                 global = get_global_options();
1773                 json_add_job_opts(root, "global options", &global->opt_list, false);
1774                 array = json_create_array();
1775                 json_object_add_value_array(root, "jobs", array);
1776         }
1777
1778         if (is_backend)
1779                 fio_server_send_job_options(&get_global_options()->opt_list, -1U);
1780
1781         for (i = 0; i < nr_ts; i++) {
1782                 ts = &threadstats[i];
1783                 rs = &runstats[ts->groupid];
1784
1785                 if (is_backend) {
1786                         fio_server_send_job_options(opt_lists[i], i);
1787                         fio_server_send_ts(ts, rs);
1788                 } else {
1789                         if (output_format & FIO_OUTPUT_TERSE)
1790                                 show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
1791                         if (output_format & FIO_OUTPUT_JSON) {
1792                                 struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
1793                                 json_array_add_value_object(array, tmp);
1794                         }
1795                         if (output_format & FIO_OUTPUT_NORMAL)
1796                                 show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
1797                 }
1798         }
1799         if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
1800                 /* disk util stats, if any */
1801                 show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
1802
1803                 show_idle_prof_stats(FIO_OUTPUT_JSON, root, &output[__FIO_OUTPUT_JSON]);
1804
1805                 json_print_object(root, &output[__FIO_OUTPUT_JSON]);
1806                 log_buf(&output[__FIO_OUTPUT_JSON], "\n");
1807                 json_free_object(root);
1808         }
1809
1810         for (i = 0; i < groupid + 1; i++) {
1811                 rs = &runstats[i];
1812
1813                 rs->groupid = i;
1814                 if (is_backend)
1815                         fio_server_send_gs(rs);
1816                 else if (output_format & FIO_OUTPUT_NORMAL)
1817                         show_group_stats(rs, &output[__FIO_OUTPUT_NORMAL]);
1818         }
1819
1820         if (is_backend)
1821                 fio_server_send_du();
1822         else if (output_format & FIO_OUTPUT_NORMAL) {
1823                 show_disk_util(0, NULL, &output[__FIO_OUTPUT_NORMAL]);
1824                 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
1825         }
1826
1827         for (i = 0; i < FIO_OUTPUT_NR; i++) {
1828                 buf_output_flush(&output[i]);
1829                 buf_output_free(&output[i]);
1830         }
1831
1832         log_info_flush();
1833         free(runstats);
1834         free(threadstats);
1835         free(opt_lists);
1836 }
1837
1838 void show_run_stats(void)
1839 {
1840         fio_mutex_down(stat_mutex);
1841         __show_run_stats();
1842         fio_mutex_up(stat_mutex);
1843 }
1844
1845 void __show_running_run_stats(void)
1846 {
1847         struct thread_data *td;
1848         unsigned long long *rt;
1849         struct timeval tv;
1850         int i;
1851
1852         fio_mutex_down(stat_mutex);
1853
1854         rt = malloc(thread_number * sizeof(unsigned long long));
1855         fio_gettime(&tv, NULL);
1856
1857         for_each_td(td, i) {
1858                 td->update_rusage = 1;
1859                 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1860                 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1861                 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1862                 td->ts.total_run_time = mtime_since(&td->epoch, &tv);
1863
1864                 rt[i] = mtime_since(&td->start, &tv);
1865                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1866                         td->ts.runtime[DDIR_READ] += rt[i];
1867                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1868                         td->ts.runtime[DDIR_WRITE] += rt[i];
1869                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1870                         td->ts.runtime[DDIR_TRIM] += rt[i];
1871         }
1872
1873         for_each_td(td, i) {
1874                 if (td->runstate >= TD_EXITED)
1875                         continue;
1876                 if (td->rusage_sem) {
1877                         td->update_rusage = 1;
1878                         fio_mutex_down(td->rusage_sem);
1879                 }
1880                 td->update_rusage = 0;
1881         }
1882
1883         __show_run_stats();
1884
1885         for_each_td(td, i) {
1886                 if (td_read(td) && td->ts.io_bytes[DDIR_READ])
1887                         td->ts.runtime[DDIR_READ] -= rt[i];
1888                 if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
1889                         td->ts.runtime[DDIR_WRITE] -= rt[i];
1890                 if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
1891                         td->ts.runtime[DDIR_TRIM] -= rt[i];
1892         }
1893
1894         free(rt);
1895         fio_mutex_up(stat_mutex);
1896 }
1897
1898 static int status_interval_init;
1899 static struct timeval status_time;
1900 static int status_file_disabled;
1901
1902 #define FIO_STATUS_FILE         "fio-dump-status"
1903
1904 static int check_status_file(void)
1905 {
1906         struct stat sb;
1907         const char *temp_dir;
1908         char fio_status_file_path[PATH_MAX];
1909
1910         if (status_file_disabled)
1911                 return 0;
1912
1913         temp_dir = getenv("TMPDIR");
1914         if (temp_dir == NULL) {
1915                 temp_dir = getenv("TEMP");
1916                 if (temp_dir && strlen(temp_dir) >= PATH_MAX)
1917                         temp_dir = NULL;
1918         }
1919         if (temp_dir == NULL)
1920                 temp_dir = "/tmp";
1921
1922         snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
1923
1924         if (stat(fio_status_file_path, &sb))
1925                 return 0;
1926
1927         if (unlink(fio_status_file_path) < 0) {
1928                 log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
1929                                                         strerror(errno));
1930                 log_err("fio: disabling status file updates\n");
1931                 status_file_disabled = 1;
1932         }
1933
1934         return 1;
1935 }
1936
1937 void check_for_running_stats(void)
1938 {
1939         if (status_interval) {
1940                 if (!status_interval_init) {
1941                         fio_gettime(&status_time, NULL);
1942                         status_interval_init = 1;
1943                 } else if (mtime_since_now(&status_time) >= status_interval) {
1944                         show_running_run_stats();
1945                         fio_gettime(&status_time, NULL);
1946                         return;
1947                 }
1948         }
1949         if (check_status_file()) {
1950                 show_running_run_stats();
1951                 return;
1952         }
1953 }
1954
1955 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
1956 {
1957         double val = data;
1958         double delta;
1959
1960         if (data > is->max_val)
1961                 is->max_val = data;
1962         if (data < is->min_val)
1963                 is->min_val = data;
1964
1965         delta = val - is->mean.u.f;
1966         if (delta) {
1967                 is->mean.u.f += delta / (is->samples + 1.0);
1968                 is->S.u.f += delta * (val - is->mean.u.f);
1969         }
1970
1971         is->samples++;
1972 }
1973
1974 /*
1975  * Return a struct io_logs, which is added to the tail of the log
1976  * list for 'iolog'.
1977  */
1978 static struct io_logs *get_new_log(struct io_log *iolog)
1979 {
1980         size_t new_size, new_samples;
1981         struct io_logs *cur_log;
1982
1983         /*
1984          * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
1985          * forever
1986          */
1987         if (!iolog->cur_log_max)
1988                 new_samples = DEF_LOG_ENTRIES;
1989         else {
1990                 new_samples = iolog->cur_log_max * 2;
1991                 if (new_samples > MAX_LOG_ENTRIES)
1992                         new_samples = MAX_LOG_ENTRIES;
1993         }
1994
1995         new_size = new_samples * log_entry_sz(iolog);
1996
1997         cur_log = smalloc(sizeof(*cur_log));
1998         if (cur_log) {
1999                 INIT_FLIST_HEAD(&cur_log->list);
2000                 cur_log->log = malloc(new_size);
2001                 if (cur_log->log) {
2002                         cur_log->nr_samples = 0;
2003                         cur_log->max_samples = new_samples;
2004                         flist_add_tail(&cur_log->list, &iolog->io_logs);
2005                         iolog->cur_log_max = new_samples;
2006                         return cur_log;
2007                 }
2008                 sfree(cur_log);
2009         }
2010
2011         return NULL;
2012 }
2013
2014 /*
2015  * Add and return a new log chunk, or return current log if big enough
2016  */
2017 static struct io_logs *regrow_log(struct io_log *iolog)
2018 {
2019         struct io_logs *cur_log;
2020         int i;
2021
2022         if (!iolog || iolog->disabled)
2023                 goto disable;
2024
2025         cur_log = iolog_cur_log(iolog);
2026         if (!cur_log) {
2027                 cur_log = get_new_log(iolog);
2028                 if (!cur_log)
2029                         return NULL;
2030         }
2031
2032         if (cur_log->nr_samples < cur_log->max_samples)
2033                 return cur_log;
2034
2035         /*
2036          * No room for a new sample. If we're compressing on the fly, flush
2037          * out the current chunk
2038          */
2039         if (iolog->log_gz) {
2040                 if (iolog_cur_flush(iolog, cur_log)) {
2041                         log_err("fio: failed flushing iolog! Will stop logging.\n");
2042                         return NULL;
2043                 }
2044         }
2045
2046         /*
2047          * Get a new log array, and add to our list
2048          */
2049         cur_log = get_new_log(iolog);
2050         if (!cur_log) {
2051                 log_err("fio: failed extending iolog! Will stop logging.\n");
2052                 return NULL;
2053         }
2054
2055         if (!iolog->pending || !iolog->pending->nr_samples)
2056                 return cur_log;
2057
2058         /*
2059          * Flush pending items to new log
2060          */
2061         for (i = 0; i < iolog->pending->nr_samples; i++) {
2062                 struct io_sample *src, *dst;
2063
2064                 src = get_sample(iolog, iolog->pending, i);
2065                 dst = get_sample(iolog, cur_log, i);
2066                 memcpy(dst, src, log_entry_sz(iolog));
2067         }
2068         cur_log->nr_samples = iolog->pending->nr_samples;
2069
2070         iolog->pending->nr_samples = 0;
2071         return cur_log;
2072 disable:
2073         if (iolog)
2074                 iolog->disabled = true;
2075         return NULL;
2076 }
2077
2078 void regrow_logs(struct thread_data *td)
2079 {
2080         regrow_log(td->slat_log);
2081         regrow_log(td->clat_log);
2082         regrow_log(td->clat_hist_log);
2083         regrow_log(td->lat_log);
2084         regrow_log(td->bw_log);
2085         regrow_log(td->iops_log);
2086         td->flags &= ~TD_F_REGROW_LOGS;
2087 }
2088
2089 static struct io_logs *get_cur_log(struct io_log *iolog)
2090 {
2091         struct io_logs *cur_log;
2092
2093         cur_log = iolog_cur_log(iolog);
2094         if (!cur_log) {
2095                 cur_log = get_new_log(iolog);
2096                 if (!cur_log)
2097                         return NULL;
2098         }
2099
2100         if (cur_log->nr_samples < cur_log->max_samples)
2101                 return cur_log;
2102
2103         /*
2104          * Out of space. If we're in IO offload mode, or we're not doing
2105          * per unit logging (hence logging happens outside of the IO thread
2106          * as well), add a new log chunk inline. If we're doing inline
2107          * submissions, flag 'td' as needing a log regrow and we'll take
2108          * care of it on the submission side.
2109          */
2110         if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
2111             !per_unit_log(iolog))
2112                 return regrow_log(iolog);
2113
2114         iolog->td->flags |= TD_F_REGROW_LOGS;
2115         assert(iolog->pending->nr_samples < iolog->pending->max_samples);
2116         return iolog->pending;
2117 }
2118
2119 static void __add_log_sample(struct io_log *iolog, union io_sample_data data,
2120                              enum fio_ddir ddir, unsigned int bs,
2121                              unsigned long t, uint64_t offset)
2122 {
2123         struct io_logs *cur_log;
2124
2125         if (iolog->disabled)
2126                 return;
2127         if (flist_empty(&iolog->io_logs))
2128                 iolog->avg_last = t;
2129
2130         cur_log = get_cur_log(iolog);
2131         if (cur_log) {
2132                 struct io_sample *s;
2133
2134                 s = get_sample(iolog, cur_log, cur_log->nr_samples);
2135
2136                 s->data = data;
2137                 s->time = t + (iolog->td ? iolog->td->unix_epoch : 0);
2138                 io_sample_set_ddir(iolog, s, ddir);
2139                 s->bs = bs;
2140
2141                 if (iolog->log_offset) {
2142                         struct io_sample_offset *so = (void *) s;
2143
2144                         so->offset = offset;
2145                 }
2146
2147                 cur_log->nr_samples++;
2148                 return;
2149         }
2150
2151         iolog->disabled = true;
2152 }
2153
2154 static inline void reset_io_stat(struct io_stat *ios)
2155 {
2156         ios->max_val = ios->min_val = ios->samples = 0;
2157         ios->mean.u.f = ios->S.u.f = 0;
2158 }
2159
2160 void reset_io_stats(struct thread_data *td)
2161 {
2162         struct thread_stat *ts = &td->ts;
2163         int i, j;
2164
2165         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2166                 reset_io_stat(&ts->clat_stat[i]);
2167                 reset_io_stat(&ts->slat_stat[i]);
2168                 reset_io_stat(&ts->lat_stat[i]);
2169                 reset_io_stat(&ts->bw_stat[i]);
2170                 reset_io_stat(&ts->iops_stat[i]);
2171
2172                 ts->io_bytes[i] = 0;
2173                 ts->runtime[i] = 0;
2174                 ts->total_io_u[i] = 0;
2175                 ts->short_io_u[i] = 0;
2176                 ts->drop_io_u[i] = 0;
2177
2178                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
2179                         ts->io_u_plat[i][j] = 0;
2180         }
2181
2182         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
2183                 ts->io_u_map[i] = 0;
2184                 ts->io_u_submit[i] = 0;
2185                 ts->io_u_complete[i] = 0;
2186         }
2187
2188         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
2189                 ts->io_u_lat_u[i] = 0;
2190         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
2191                 ts->io_u_lat_m[i] = 0;
2192
2193         ts->total_submit = 0;
2194         ts->total_complete = 0;
2195 }
2196
2197 static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
2198                               unsigned long elapsed, bool log_max)
2199 {
2200         /*
2201          * Note an entry in the log. Use the mean from the logged samples,
2202          * making sure to properly round up. Only write a log entry if we
2203          * had actual samples done.
2204          */
2205         if (iolog->avg_window[ddir].samples) {
2206                 union io_sample_data data;
2207
2208                 if (log_max)
2209                         data.val = iolog->avg_window[ddir].max_val;
2210                 else
2211                         data.val = iolog->avg_window[ddir].mean.u.f + 0.50;
2212
2213                 __add_log_sample(iolog, data, ddir, 0, elapsed, 0);
2214         }
2215
2216         reset_io_stat(&iolog->avg_window[ddir]);
2217 }
2218
2219 static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
2220                              bool log_max)
2221 {
2222         int ddir;
2223
2224         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2225                 __add_stat_to_log(iolog, ddir, elapsed, log_max);
2226 }
2227
2228 static long add_log_sample(struct thread_data *td, struct io_log *iolog,
2229                            union io_sample_data data, enum fio_ddir ddir,
2230                            unsigned int bs, uint64_t offset)
2231 {
2232         unsigned long elapsed, this_window;
2233
2234         if (!ddir_rw(ddir))
2235                 return 0;
2236
2237         elapsed = mtime_since_now(&td->epoch);
2238
2239         /*
2240          * If no time averaging, just add the log sample.
2241          */
2242         if (!iolog->avg_msec) {
2243                 __add_log_sample(iolog, data, ddir, bs, elapsed, offset);
2244                 return 0;
2245         }
2246
2247         /*
2248          * Add the sample. If the time period has passed, then
2249          * add that entry to the log and clear.
2250          */
2251         add_stat_sample(&iolog->avg_window[ddir], data.val);
2252
2253         /*
2254          * If period hasn't passed, adding the above sample is all we
2255          * need to do.
2256          */
2257         this_window = elapsed - iolog->avg_last;
2258         if (elapsed < iolog->avg_last)
2259                 return iolog->avg_last - elapsed;
2260         else if (this_window < iolog->avg_msec) {
2261                 int diff = iolog->avg_msec - this_window;
2262
2263                 if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
2264                         return diff;
2265         }
2266
2267         _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
2268
2269         iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
2270         return iolog->avg_msec;
2271 }
2272
2273 void finalize_logs(struct thread_data *td, bool unit_logs)
2274 {
2275         unsigned long elapsed;
2276
2277         elapsed = mtime_since_now(&td->epoch);
2278
2279         if (td->clat_log && unit_logs)
2280                 _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
2281         if (td->slat_log && unit_logs)
2282                 _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
2283         if (td->lat_log && unit_logs)
2284                 _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
2285         if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
2286                 _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
2287         if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
2288                 _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
2289 }
2290
2291 void add_agg_sample(union io_sample_data data, enum fio_ddir ddir, unsigned int bs)
2292 {
2293         struct io_log *iolog;
2294
2295         if (!ddir_rw(ddir))
2296                 return;
2297
2298         iolog = agg_io_log[ddir];
2299         __add_log_sample(iolog, data, ddir, bs, mtime_since_genesis(), 0);
2300 }
2301
2302 static void add_clat_percentile_sample(struct thread_stat *ts,
2303                                 unsigned long usec, enum fio_ddir ddir)
2304 {
2305         unsigned int idx = plat_val_to_idx(usec);
2306         assert(idx < FIO_IO_U_PLAT_NR);
2307
2308         ts->io_u_plat[ddir][idx]++;
2309 }
2310
2311 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
2312                      unsigned long usec, unsigned int bs, uint64_t offset)
2313 {
2314         unsigned long elapsed, this_window;
2315         struct thread_stat *ts = &td->ts;
2316         struct io_log *iolog = td->clat_hist_log;
2317
2318         td_io_u_lock(td);
2319
2320         add_stat_sample(&ts->clat_stat[ddir], usec);
2321
2322         if (td->clat_log)
2323                 add_log_sample(td, td->clat_log, sample_val(usec), ddir, bs,
2324                                offset);
2325
2326         if (ts->clat_percentiles)
2327                 add_clat_percentile_sample(ts, usec, ddir);
2328
2329         if (iolog && iolog->hist_msec) {
2330                 struct io_hist *hw = &iolog->hist_window[ddir];
2331
2332                 hw->samples++;
2333                 elapsed = mtime_since_now(&td->epoch);
2334                 if (!hw->hist_last)
2335                         hw->hist_last = elapsed;
2336                 this_window = elapsed - hw->hist_last;
2337
2338                 if (this_window >= iolog->hist_msec) {
2339                         unsigned int *io_u_plat;
2340                         struct io_u_plat_entry *dst;
2341
2342                         /*
2343                          * Make a byte-for-byte copy of the latency histogram
2344                          * stored in td->ts.io_u_plat[ddir], recording it in a
2345                          * log sample. Note that the matching call to free() is
2346                          * located in iolog.c after printing this sample to the
2347                          * log file.
2348                          */
2349                         io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
2350                         dst = malloc(sizeof(struct io_u_plat_entry));
2351                         memcpy(&(dst->io_u_plat), io_u_plat,
2352                                 FIO_IO_U_PLAT_NR * sizeof(unsigned int));
2353                         flist_add(&dst->list, &hw->list);
2354                         __add_log_sample(iolog, sample_plat(dst), ddir, bs,
2355                                                 elapsed, offset);
2356
2357                         /*
2358                          * Update the last time we recorded as being now, minus
2359                          * any drift in time we encountered before actually
2360                          * making the record.
2361                          */
2362                         hw->hist_last = elapsed - (this_window - iolog->hist_msec);
2363                         hw->samples = 0;
2364                 }
2365         }
2366
2367         td_io_u_unlock(td);
2368 }
2369
2370 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
2371                      unsigned long usec, unsigned int bs, uint64_t offset)
2372 {
2373         struct thread_stat *ts = &td->ts;
2374
2375         if (!ddir_rw(ddir))
2376                 return;
2377
2378         td_io_u_lock(td);
2379
2380         add_stat_sample(&ts->slat_stat[ddir], usec);
2381
2382         if (td->slat_log)
2383                 add_log_sample(td, td->slat_log, sample_val(usec), ddir, bs, offset);
2384
2385         td_io_u_unlock(td);
2386 }
2387
2388 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
2389                     unsigned long usec, unsigned int bs, uint64_t offset)
2390 {
2391         struct thread_stat *ts = &td->ts;
2392
2393         if (!ddir_rw(ddir))
2394                 return;
2395
2396         td_io_u_lock(td);
2397
2398         add_stat_sample(&ts->lat_stat[ddir], usec);
2399
2400         if (td->lat_log)
2401                 add_log_sample(td, td->lat_log, sample_val(usec), ddir, bs,
2402                                offset);
2403
2404         td_io_u_unlock(td);
2405 }
2406
2407 void add_bw_sample(struct thread_data *td, struct io_u *io_u,
2408                    unsigned int bytes, unsigned long spent)
2409 {
2410         struct thread_stat *ts = &td->ts;
2411         unsigned long rate;
2412
2413         if (spent)
2414                 rate = bytes * 1000 / spent;
2415         else
2416                 rate = 0;
2417
2418         td_io_u_lock(td);
2419
2420         add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
2421
2422         if (td->bw_log)
2423                 add_log_sample(td, td->bw_log, sample_val(rate), io_u->ddir,
2424                                bytes, io_u->offset);
2425
2426         td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
2427         td_io_u_unlock(td);
2428 }
2429
2430 static int __add_samples(struct thread_data *td, struct timeval *parent_tv,
2431                          struct timeval *t, unsigned int avg_time,
2432                          uint64_t *this_io_bytes, uint64_t *stat_io_bytes,
2433                          struct io_stat *stat, struct io_log *log,
2434                          bool is_kb)
2435 {
2436         unsigned long spent, rate;
2437         enum fio_ddir ddir;
2438         unsigned int next, next_log;
2439
2440         next_log = avg_time;
2441
2442         spent = mtime_since(parent_tv, t);
2443         if (spent < avg_time && avg_time - spent >= LOG_MSEC_SLACK)
2444                 return avg_time - spent;
2445
2446         td_io_u_lock(td);
2447
2448         /*
2449          * Compute both read and write rates for the interval.
2450          */
2451         for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
2452                 uint64_t delta;
2453
2454                 delta = this_io_bytes[ddir] - stat_io_bytes[ddir];
2455                 if (!delta)
2456                         continue; /* No entries for interval */
2457
2458                 if (spent) {
2459                         if (is_kb)
2460                                 rate = delta * 1000 / spent / 1024; /* KiB/s */
2461                         else
2462                                 rate = (delta * 1000) / spent;
2463                 } else
2464                         rate = 0;
2465
2466                 add_stat_sample(&stat[ddir], rate);
2467
2468                 if (log) {
2469                         unsigned int bs = 0;
2470
2471                         if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
2472                                 bs = td->o.min_bs[ddir];
2473
2474                         next = add_log_sample(td, log, sample_val(rate), ddir, bs, 0);
2475                         next_log = min(next_log, next);
2476                 }
2477
2478                 stat_io_bytes[ddir] = this_io_bytes[ddir];
2479         }
2480
2481         timeval_add_msec(parent_tv, avg_time);
2482
2483         td_io_u_unlock(td);
2484
2485         if (spent <= avg_time)
2486                 next = avg_time;
2487         else
2488                 next = avg_time - (1 + spent - avg_time);
2489
2490         return min(next, next_log);
2491 }
2492
2493 static int add_bw_samples(struct thread_data *td, struct timeval *t)
2494 {
2495         return __add_samples(td, &td->bw_sample_time, t, td->o.bw_avg_time,
2496                                 td->this_io_bytes, td->stat_io_bytes,
2497                                 td->ts.bw_stat, td->bw_log, true);
2498 }
2499
2500 void add_iops_sample(struct thread_data *td, struct io_u *io_u,
2501                      unsigned int bytes)
2502 {
2503         struct thread_stat *ts = &td->ts;
2504
2505         td_io_u_lock(td);
2506
2507         add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
2508
2509         if (td->iops_log)
2510                 add_log_sample(td, td->iops_log, sample_val(1), io_u->ddir,
2511                                bytes, io_u->offset);
2512
2513         td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
2514         td_io_u_unlock(td);
2515 }
2516
2517 static int add_iops_samples(struct thread_data *td, struct timeval *t)
2518 {
2519         return __add_samples(td, &td->iops_sample_time, t, td->o.iops_avg_time,
2520                                 td->this_io_blocks, td->stat_io_blocks,
2521                                 td->ts.iops_stat, td->iops_log, false);
2522 }
2523
2524 /*
2525  * Returns msecs to next event
2526  */
2527 int calc_log_samples(void)
2528 {
2529         struct thread_data *td;
2530         unsigned int next = ~0U, tmp;
2531         struct timeval now;
2532         int i;
2533
2534         fio_gettime(&now, NULL);
2535
2536         for_each_td(td, i) {
2537                 if (!td->o.stats)
2538                         continue;
2539                 if (in_ramp_time(td) ||
2540                     !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
2541                         next = min(td->o.iops_avg_time, td->o.bw_avg_time);
2542                         continue;
2543                 }
2544                 if (!td->bw_log ||
2545                         (td->bw_log && !per_unit_log(td->bw_log))) {
2546                         tmp = add_bw_samples(td, &now);
2547                         if (tmp < next)
2548                                 next = tmp;
2549                 }
2550                 if (!td->iops_log ||
2551                         (td->iops_log && !per_unit_log(td->iops_log))) {
2552                         tmp = add_iops_samples(td, &now);
2553                         if (tmp < next)
2554                                 next = tmp;
2555                 }
2556         }
2557
2558         return next == ~0U ? 0 : next;
2559 }
2560
2561 void stat_init(void)
2562 {
2563         stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2564 }
2565
2566 void stat_exit(void)
2567 {
2568         /*
2569          * When we have the mutex, we know out-of-band access to it
2570          * have ended.
2571          */
2572         fio_mutex_down(stat_mutex);
2573         fio_mutex_remove(stat_mutex);
2574 }
2575
2576 /*
2577  * Called from signal handler. Wake up status thread.
2578  */
2579 void show_running_run_stats(void)
2580 {
2581         helper_do_stat();
2582 }
2583
2584 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
2585 {
2586         /* Ignore io_u's which span multiple blocks--they will just get
2587          * inaccurate counts. */
2588         int idx = (io_u->offset - io_u->file->file_offset)
2589                         / td->o.bs[DDIR_TRIM];
2590         uint32_t *info = &td->ts.block_infos[idx];
2591         assert(idx < td->ts.nr_block_infos);
2592         return info;
2593 }